CCSIT GRADUATION PROJECTS PROCEEDING 2022

CCSIT GRADUATION

PROJECTS PROCEEDING 2022

2

INTRODUCTION

، تعمل كلية علوم الحاسب وتقنية المعلومات بجامعة اإلمام عبد الرحمن بن فيصل عىل تشجيع ودعم البحث العلميي المستوى العاشر عىل النشر العلمي من خالل ومن هنا شجعت طالب وطالبات كلية علوم الحاسب وتقنية المعلومات ف

ي الكلية وسابقة تفتخر فيها الكلية، فنادرا ي مشاري ع التخرج وما توصلوا اليه من نتائج. كانت هذه تجربة جديدة ف

عملهم ف ي اوعية نشر معتمدة. نتج عن هذه التجربة ان استطاع طالب وطالبات الكلية لبةان ينشر ط

البكالوريوس بحث علمي ف ي هذا الكتيب تجدون األوراق العلمية المنشورة

ي مجالت او مؤتمرات مرموقة. ف نشر عدد من األبحاث الجديرة باالهتمام ف

. 2022/ 2021من طلبة الكلية لعام

The College of Computer Science and Information Technology (CCSIT) at Imam Abdulrahman

Bin Faisal University (IAU) encourages and supports scientific research. As a new experience,

CCSIT encouraged senior students to publish their findings and achievements during their

works in the graduation projects. Proudly, we can announce that our students succeed to publish

their work in journals and conferences with high reputation. In this proceeding you will find

the scientific papers published by our students for the year 2021/2022.

3

CCSIT 2020 GP PUBLICATIONS

Section I: Journal publications

Journal 1: Phishing Email Detection Using Machine Learning Techniques

Journal 2: Intelligent Techniques for Predicting Stock Market Prices: A Critical Survey

Section II: Conference publications

Con1: Sa’ah: Creative Eco-Friendly Mobile Application That Encourages Living Sustainably

Con2: Aknaf Website: Interactive Website to Automate the Institution’s Work Con3: Flourish: Requirements and Design of an Android Application Prototype for Various

Symptoms Management in ADHD Patients

Con4: Machine Learning Based Preemptive Diagnosis of Lung Cancer Using Clinical Data

Con5: Leen: Web-based Platform for Pet Adoption

Con6: Road Damages Detection and Classification using Deep Learning and UAVs

Con7: a comparison between vgg16 and xception models used as encoders for image

captioning

Con8: Smart Inventory System

Section III: Others

Intelligent Watering System

4

JOURNAL PAPERS

Proceedings of Graduation Project Showcase 2022

1 | P a g e

Published In: IJCSNS International Journal of Computer Science and Network Security, VOL.22 No.3, March 2022

https://doi.org/10.22937/IJCSNS.2020.22.3.1

Phishing Email Detection Using Machine Learning Techniques

Hussain Alattas1, Fay Aljohar2, Hawra Aljunibi3, Muneera Alweheibi4, Rawan Alrashdi5, Ghadeer Al

azman6, Abdulrahman Alharby7 and Naya Nagy8

[email protected] [email protected] [email protected] [email protected] 2170002618

@iau.edu.sa [email protected] [email protected] [email protected] University of Imam Abdulrahman bin Faisal, College of Computer Science, and Information Technology, KSA

Abstract Phishing is a social engineering technique that mainly aims to steal

personal or confidential data and may harm the target individual

or organization in many ways. In phishing, fraudsters hide their

identity as legitimate people, banks, or institutions, whether

governmental or private. And since e-mail communication is the

most used method in transmitting confidential or official messages,

fraudsters normally target the email users to send their deceptive

messages in order to extract data. However, this paper presents an

overview of previously conducted studies with respect to detecting

phishing email messages using machine learning. The paper’s

objective is to analyze and assess the procedures of previously

proposed models, datasets, and their results within the specified

scope.

Keywords: Phishing Attacks, Machine Learning, Phishing Emails, Social

Engineering, Email Security.

1. Introduction

Phishing emails represent a threat in the world of the

Internet, as email is the main place to send messages,

whether personally or officially, as many individuals

depend on it and review it daily. The interaction of one

individual in an organization with a phishing message

may lead to the destruction of the entire organization,

this is what we mean by a threat phishing message. In

this paper, we discuss some of the previous research

on detecting phishing attacks in email and some

models and suggested features in detecting these

attacks. We also present a comparative study of classic

machine learning techniques such as Random Forest,

Random Forest, Naive Bayes, Decision Tree, and

Support Vector Machine (SVM). This paper is

sectioned by a problem statement, background, review

of literature that has three sub-sections supervised

machine learning techniques, non-supervised, and

others; moreover, it illustrates a comparison table

between models in the aspect of approaches,

limitations, algorithms, response time, and accuracy.

2. Problem Statement

A phishing attack is generally accomplished by

sending email messages that appear to come from a

trusted source and require the user to enter financial,

personal, or confidential data. The problem is when

the user interacts with the email and sends the

requested response, either by replying to the email by

sending confidential data, visiting a website, or

clicking on a link. Attackers are always coming up

with new and inventive ways to dupe people into

thinking their activities are related to a legitimate

website or email. The user interacts without thinking

when the situation seems to be dangerous, fearful,

urgent, etc. Most end users usually make the decision

based on how they look and feel.

3. Background

In the early 1990s, a huge number of users with false

credit card details created an algorithm for stealing

user information, they registered themselves on

America Online (AOL) site without any validation and

started using system resources. When AOL eliminated

the random credit card generators in 1995, the Warez

group shifted to other techniques, including

communicating with individuals via AOL Messenger

while pretending to be AOL employees and requesting

their personal information. In 1996, American On

line's Usenet group posted the first mention of the term

"phishing" in response [1]. Phishing occurs when

cybercriminals send malicious emails to trick a victim

into falling for a scam. The goal is usually to persuade

users to divulge sensitive information such as financial

data or system credentials. The advantages of phishing

for cybercriminals include its simplicity, low cost, and

mailto:[email protected]








2 | P a g e



effectiveness. Attackers can easily gain access to

valuable information with very little effort and for a

low price. Due to this, we are going to discuss a variety

of machine learning models to detect such phishing e-

mails and then block them [1]. Machine learning is a

method of analyzing data that automates the process

of constructing analytical models. This branch of

artificial intelligence relies on the idea that computers

can identify patterns, learn from data, and make

decisions without the need for any human interference

[2].

4. Review of literature

As phishing emails constitute the primary gateway to

phishing websites, several papers were examined that

discuss phishing email detection and classification

techniques. A major approach for phishing email

detection and classification is to employ machine

learning techniques.

4.1 Machine Learning and Phishing Emails

Detection

Machine learning is a critical ally in fighting phishing

emails. Mostly, it investigates the content, metadata,

context, and regular user behavior to analyze and

detect phishing. Machine-learning includes several

types such as supervised machine learning which

utilizes label data to train models, and unsupervised

machine learning which utilizes patterns from

unlabeled data to train them. Though, unsupervised

machine learning may give less accurate results

compared to supervised machine learning [3].

Examples of previous work regarding these machine

learning techniques are going to be discussed in the

subsequent sections.

4.1.1 Supervised Machine Learning Techniques

As described in [4], A. Shaheen et al. proposed a

model based on supervised machine learning

algorithms to classify phished and ham mail. In

supervised learning algorithms, a training set is used

to classify test sets. The dataset consists of 1605

emails, 1191 are ham and 414 are phished. Ham

emails are derived from a publicly available dataset,

while phished emails are derived from multiple

sources. After preprocessing and converting the

dataset, features were extracted and used to feed the

classifiers. The features are extracted from the dataset

using the Python programming language and the

Nerve Learning Toolkit. The dataset consists of

extracted features is segmented and fed into five

classifiers: Logistic, Random Forest, SVM, Voted

Perceptron, and Naive Bayes. Results showed that the

classification of emails through SVM and Random

Forest classifiers was highly accurate, achieving the

highest accuracy of 99.8%.

Akash Junnarkar et al. [5] built a comprehensive

system for spam classification using semantics-based

text classification and URL-based filtering. They

establish a spam classification system that followed a

two-step methodology to ensure that all mail received

was either spam or not. The process begins with text

classification and is followed by URL analysis and

filtering to determine whether any links present in the

email are malicious. Five machine learning algorithms

were considered for text classification: K-Nearest

Neighbours, Naive Bayes, Decision Tree, Random

Forest, and SVM. The highest accuracy is obtained

with Naive Bayes and SVM, hitting a 97.83 %

accuracy rate for SVM and 95.48 % for Naive Bayes.

As Naive Bayes and S had the highest accuracy, they

were implemented in the final model to identify

trigger words within the text. Lists of spam trigger

words and blacklisted URLs were compiled using

several datasets. The model was hosted as an API that

was called by JavaScript code in Google Apps script

to process emails in real-time.

In [6], Jameel et al. proposed a phishing detection

model that uses a feed-forward neural network. The

model was created based on the characteristics of

phishing emails. Thus, a set of 18 features were

extracted from the tested email, these email features

appear in the header and the HTML body of the email.

In a subsequent step, a multilayer feedforward neural

network is used to classify the tested email into

phishing or ham email. A total of 9100 phishing and

ham emails have been used to test this model; 4550 of

these emails are phishing emails were collected from

publicly available phishing Corpus

(www.monkey.org), while 4550 of these samples are

ham emails were collected from the Spam Assassin

project's ham corpora. According to the testing results,


3 | P a g e



the identification rate of this model was excellent

(98.7%).

A method based on neural networks was proposed by

George et al. [7]. The team used two datasets

consisting of 4500 emails phish and ham. To identify

ham and phish emails, they applied various

algorithms, including Feedforward Neural Network

(FNN) with back propagation, and fist order statistical

measures. As a result, the false-negative rate and the

false positive rate are exceptionally low. With 12

features, 99.95% of the results were classified

correctly.

Kumar et al. [8] investigated the detection of phishing

emails lacking links and URLs. In their proposed

work, they have used NLP and WordNet. Using 600

phishing emails and 400 legitimate emails, they have

compiled a list of features including the absence of

recipients' names, asking for money, or mentioning

money, a sense of urgency, and a sense of urgency that

lures victims to respond. They had based their work

on Stanford Core NLP's application program interface

to identify all the words found in phishing emails.

Harikrishnan et al. proposed [9] (Term Frequency

Inverse Document Frequency) TFID+ (Singular Value

Decomposition) SVD and TFIDF+ (Nonnegative

Matrix Factorization) NMF to evaluate if it is in fact

phishing email or not. The model starts by using email

datasets with and without headers passed to data pre-

processing. Then, to convert words to a numeric

representation it uses TFIDF. After that, it uses SVD

and NMF to extract features. Lastly, to decide whether

it is legitimate or not, classical Machine Learning

(ML) techniques are utilized. The accuracy of the

result for this model was low due to the highly

imbalanced dataset.

Senturkurk et al. [10] proposed a model that begins

with data set training by concentrating on the email's

body and ignoring the attachments and header. After

the data sets are ready, it starts the feature selection.

Then passed it to Waikato Environment for

Knowledge Analysis (WEKA) tool after converting it

to the proper format. Later, a sub-list is initiated below

this new decision node and a sub-decision tree is built.

After that, a different algorithm used: Naïve Bayes and

decision tree. Finally, the result shows it will appear

high accuracy rate when a supplied test is selected and

performing datasets for all operations is in a real-time

environment.

The proposed approach by Hamid et al. [11] is called

the Hybrid Feature Selection (HFS). HFS applies to

6923 datasets from both Nazario and SpamAssassin

datasets. In addition, it analyzes the sender behavior to

resolve a feature matrix utilizing seven email relevant

features to determine whether an email is phishing or

not. Further, in order for HFS to classify the email, it

uses an algorithm named Bayes Net algorithm for

email classifications.

As shown by Adewumi and Akinyelu [12] the Firefly

Algorithm (FFA) is combined with the (SVM) for

machine learning classification to build a hybrid

classifier called FFA_SVM. For the purpose of

evaluating the FFA_SVM algorithm, a database was

constructed of 4000 phishing and ham emails along

with their features. FFA_SVM has outperformed the

standard SVM.

Alayham et al. [13] design and develop a tool that

detects the source code of a phishing site associated

with a Gmail account using a decision tree algorithm

and generates a report of phishing sites attached to a

victim's email as the percentages of phishing emails

stored in the user's mailbox. Also, the application can

send notifications to the user regarding a phishing site

that was detected in the incoming message. The Agile

Unified Process (AUP) methodology was used to

implement the tool.

Husak and J. Cegan [14] Develop an automated tool

to deal with PhiGARo phishing incidents that identify

individuals who respond to phishing attack attempts.

The network traffic of the honeypot is monitored, and

any phishing emails detected are sent to the PhiGARo

tool. The PhiGARo framework is divided into two

parts, the Phishing Incident Handling section and the

Phishing Response and Detection section. Initially, the

phishing incident is reported by the user who

recognizes the phishing message in their mailbox.

PhiGARo is implemented by Incident Handler

manually, then interpreting the results, blocking the

phishing email or URL, and finally notifying the

victims.


4 | P a g e



Egozi and Verma [15] created a phishing email

detection tool with 26 features. Features include word

count, stop words, repeating punctuation, and unique

words. 17 machine languages were studied and

categorized under weighted and unweighted, based on

the results, the weighted linear SVM algorithm

represented the best model.

Unnithan et al. [16] proposed a model based on a

variety of mathematical algorithms to measure if an

email is a legitimate email or not. Consists of two

dataset emails with headers and without headers. This

sample is sent to count-based representation Term

Frequency Inverse Document Frequency (TFIDF) and

then combined with domain-level features to convert

the input to an understandable input for machine

learning algorithms. The last step in the model to

decide whether it is a legitimate or phishing email is

passed to several machine learning such as logistic

regression, Naive Bayes, SVM.

4.1.2 Unsupervised Machine Learning

Techniques

Fuertes et al. [17] is described how to develop a

Scrum-based algorithm implementation of automatic

learning, feature selection, and neural networks, with

the goal of attack detecting and mitigating from inside

the email server. The samples were divided into three

different time periods and tested on a different dataset

that was previously merged. Feature Selection, Neural

Networks, Agile Scrum methodology, and Matlab

process tool are used during the implementation of the

proposed algorithm. Because the developed methods

complement each other during detection, the acquired

results from the concept tests are highly promising.

The findings of the three data sets were evaluated, and

the average accuracy was 93.9%, and to validate the

results obtained the source of information from the

Phish Tank blacklist was used.

Andrade et al. [18] create a Python software that uses

a machine-learning algorithm to learn how to

recognize bad URLs, then provides relevant analysis

and information about the bad URLs. The program

also includes an examination of the analysis of

anomalous behavior linked to phishing web attacks, as

well as how machine learning techniques may be used

to counter the problem. This analysis is carried out

using tainted datasets provided by Kaggle Phishing

Dataset and Python tools to develop machine learning

to detect phishing attacks by analyzing URLs to

determine whether they are good or bad based on

specific characteristics of URLs, with the goal of

providing information in real-time so that proactive

decisions can be made to reduce the impact of the

attack. When information is added to machine

learning algorithms and the algorithm is performed,

the accuracy and error are likely to improve.

Unnithan et al. [19] proposed a model based on a

variety of mathematical algorithms to measure if an

email is a legitimate email or not. Consists of two

dataset emails with headers and without headers. This

sample is sent to count-based representation TF-IDF

and then combined with domain-level features to

convert the input to an understandable input for

machine learning algorithms. The last step in the

model to decide whether it is a legitimate or phishing

email is passed to several machine learning such as

logistic regression, Naive Bayes, Support Vector

Machine. The accuracy of this model after testing

4.1.3 Other Machine Learning Techniques

The proposed phishing detection model in [20] by

Viktorov, uses a dataset of phishing and non-phishing

emails from different websites. The model starts with

preprocessing the collected data to extract features

from each email. Second, passed to feature selection

which splits into two scenarios. Those scenarios are

automated and manually. In the manually use

clustering, which is like classification, but it is

unsupervised. third, it is passed to the classification

selection phase. fourth to multi-classifier, that uses

several algorithms to build it such as Logistic

regression, Decision Tree and Sequential minimal

optimization. The results showed that clustering will

increase the accuracy rate.

Rastenis et al. [21] discuss the Multi-Language

Spam/Phishing Classification solution that classifies

an unwanted email to either spam or phishing emails

classes through using the email body content and a

dataset that is constructed by three other known data

sets: Nazario, SpamAssassin, and VilniusTech.

Additionally, it can classify the email even if it is

written in Russian and Lithuanian languages rather


5 | P a g e



than just English through integrating with existing

classifying emails solutions and automated

translation.

Fang et al. proposed [22] an approach named THEMIS

(Greek word) that uses unbalanced dataset and divides

the email into two parts: the email’s header and body.

Then, it splits it more into two levels: the char-level,

and word-level for both header and body. Also, it

calculates the likelihood if an email is phishing by

comparing the probability with a classification value

called a threshold, if the probability is greater than this

value then it is a phishing email.

Li et al. have presented [23] the overall function of the

Long Short-Term Memory (LSTM) Network method

for big email data. LSTM cannot use an open-source

dataset; thus, a filter must be conducted manually first

of the nature of the phishing emails the enterprise

receives. After a filter has been established, both

supervised KNN and unsupervised K Means are used

to conduct labeling automation to construct a set of

samples used for phishing email detection.

5 Comparison

This section represents a comparison between given

machine learning techniques discussed in the literature

to detect phishing emails. The comparison is based on

which algorithm(s) or model(s) had been used,

accuracy, Ture Positive Rate (TPR), False Positive

Rate (FPR), datasets used, number of features,

response time, and drawbacks.

5.1 Supervised Machine Learning Techniques Comparison Table.

Author Algorithm(s) used Accuracy TPR FPR Datasets used No. of

Features Response time Drawbacks

Supervised Machine Learning Techniques

[4] Random Forest 99.87% 99.9% 0.2% N. A 9 N. A

Data used may not

reflect real life

scenarios

[5] SVM 97.83 % 53.0% 3.0%

Enron Data set

and spam.csv

Kaggle data

N. A N. A

There is no real-time

learning of email

classifiers in the

provided data sets

[6]

FNN 98.72% 98% 1.2% N. A 18

0.00000067

seconds

Increased numbers of

neurons will increase

training and testing

time

[7]

FNN

99.95% 100% 0.09% N. A 12 0.00000118

seconds N. A

[8] NLP 99.4% N. A N. A N. A N. A N. A

Unable to extract

text from email

attachment

[11] Bayes Net 94% 0.97% 0.13% Nazrario &

SpamAssassin 7 N. A

Graphical form in

phishing emails

cannot be detected

[12] SVM 99.94% N. A 0.01% Dataset consists

of 4000 emails 16 0.16 seconds N. A

Decision Tree 96.5 % 92%-

97%

PhishingCorpus

7

8.54 seconds

Dataset is highly

imbalanced Random Forest 97.1% Slow


6 | P a g e



[9]

KNN 97.6 % 8%-

26%

4.3 Seconds

Naive Base 94.7 % 0.01 seconds

AdaBoost 97.7% N. A

SVM 98.7% 0.16 seconds

Logistic Regression 96.8% 12.11 seconds

[10]

Naïve bayes

89% N. A N. A MIME 13

0.01 seconds Datasets must be

in real-time

environment to

success Decision Tree 8.54 seconds

[13] Decision Tree 95.05% N. A N. A Used 3 available

dataset 8 8.54 seconds N. A

[14] IPFIX N. A N. A Low N. A N. A 3 to 19 per day

Must support

the

trustworthiness

of honeytokens

and honeypots

[15] SVM 90% 83.0% 96.0% IWSPA 28 0.16 seconds Takes few hours to

run

[16]

Naïve Bayes 79.5% N. A N. A

Enron and

Avocado N. A

0.01 seconds

Cannot extract

feature from headers SVM 88.4%

3593/4

583

489/458

3 0.16 seconds

Logistic Regression 80.1% N. A N. A 2.11 seconds

Table 1: Supervised Machine Learning Techniques Comparison

5.2 Unsupervised Machine Learning Techniques Table.

Table 2: Unsupervised Machine Learning Techniques Comparison.

5.3 Other Machine Learning Techniques Table.



Unsupervised Machine Learning Techniques

[17] Agile Scrum 93.9% N. A 2.7% Debian

Phish Tank 7 N. A N. A

[18] Logistic 90% N. A N. A Kaggle N. A 12.11 seconds N. A

[19]

SVM 95% 3807/

3572 7/217 N. A 5 N. A

cannot extract

features from

headers

Naïve Bayes 94%

Logistic Regression 96%


7 | P a g e



Table 3: Other Machine Learning Techniques Comparison.

5.4 Analysis

According to the comparisons in Tables 1, 2, and 3,

four main models had been considered as remarkable

models among others based on different parameters

for email classification. These models are SVM, NLP,

Random Forest and Naive Bayes models. Despite the

fact that they had gained popularity in many previous

works regarding email classification techniques SVM,

NLP, Random Forest, and Naive Bayes algorithms

have very high accuracy, TPR, and FPR compared to

other algorithms with fast response times. On the other

hand, two datasets had also gained popularity in the

phishing detection field to extract informative email

features for classification, these are Spam Assassin

and Nazario corpuses. However, our literature study

had shown that there are many effective algorithms of

email classification, yet attackers are becoming more

and more sophisticated with powerful techniques.

Thus, each time ones want to decide which algorithms

or learner are best to distinguish if an email is a

phishing or non-phishing email is now becoming a

difficult challenge.

6 Conclusion

Over the past few years, the problem of phishing

emails has become more common. Phishing is a type

of attack. The intention of phishing is to obtain

personal information, such as passwords, credit card

numbers, or other account information, by using

emails. Phishing emails closely resemble legitimate

ones, making it hard for a layperson to distinguish

them. Machine learning techniques currently play a

major role in phishing email detection and

classification. Several models and approaches are

available for phishing email detection. Each approach

has its own unique advantages and capabilities, as well

as limitations. Hence, this literature review has

summarized and compared several methods and

approaches for protecting against phishing email

attacks.

References

[1] P. Verma, A. Goyal, and Y. Gigras, “Email

phishing: text classification using natural

language processing,” Comput. Sci. Inf. Technol.,

vol. 1, no. 1, pp. 1–12, 2020.

[2] E. Bisong, “What is machine learning?” in ” in

Building Machine Learning and Deep Learning



Other Machine Learning Techniques

[21]

SVM English

only

(90.07%

±3.17%)

English,

Russian and

Lithuanian

(89.2%±2.1

4)

95.2% N. A

Nazario,

SpamAssassin, and

VilniusTech.

N. A

0.16 seconds

Accuracy lessens 10%

if a mixed dataset is

used for training and

testing

Random Forest Too slow

Decision Tree 8.54 seconds

Naïve Bayes 0.01 seconds

KNN 4.3 Seconds

[22] Threshold value 99.848% 99.0% 0.043

% WordNet, Enron,

and Nazario N. A

Increased

response time N. A

[23]

KNN

95% 98% N. A Collected from a

private enterprise 7

4.3 Seconds Consume time on

constructing the filter K-Means Fast

[20]

Logistic Regression

93% N. A 4.89% Datasets consist

of 4800 emails 47

12.11 seconds Email is not clustered

before classification

which reduced the

accuracy

Decision Tree 8.54 seconds

CART N. A

SMO Medium


8 | P a g e



Models on Google Cloud Platform, Berkeley, CA:

Apress, 2019, pp. 169–170.

[3] Gatefy, “How artificial intelligence and machine

learning fight phishing,” Gatefy, 22-Mar-2021.

[Online]. Available: https://gatefy.com/blog/how-

ai-and-ml-fight-phishing/. [Accessed: 13-Mar-

2022].

[4] S. Rawal, A. Shaheen, and S. Malik, “Phishing

Detection in E-mails using Machine Learning,”

Int. J. Appl. Inf. Syst., vol. 12, no. 7, pp. 21–24,

2017.

[5] A. Junnarkar, S. Adhikari, J. Fagania, P.

Chimurkar, and D. Karia, “E-mail spam

classification via machine learning and natural

language processing,” in 2021 Third International

Conference on Intelligent Communication

Technologies and Virtual Mobile Networks

(ICICV), 2021.

[6] N. Ghazi, M. Jameel and L. E. George, “Detection

of phishing emails using feed forward neural

network,” Int. J. Comput. Appl., vol. 77, no. 7, pp.

10–15, 2013.

[7] A. A. Abdullah, L. E. George, and I. J.

Mohammed, “Research Article Email Phishing

Detection System Using Neural Network,”

Research Journal of Information Technology, vol.

6, no. 3, pp. 39–43, 2015.

[8] Aggarwal, Shivam, Vishal Kumar and Sithu D.

Sudarsan. “Identification and Detection of

Phishing Emails Using Natural Language

Processing Techniques.” SIN (2014).

[9] B, Harikrishnan & Ravi, Vinayakumar & Kp,

Soman. (2018), "A Machine Learning Approach

Towards Phishing Email Detection," CEN-

Security@IWSPA 2018.

[10] Ş. Şentürk, E. Yerli and İ. Soğukpınar, "Email

phishing detection and prevention by using data

mining techniques," 2017 International

Conference on Computer Science and

Engineering (UBMK), 2017, pp. 707-712, doi:

10.1109/UBMK.2017.8093510.

[11] I. R. A Hamid, J. Abawajy, and T.-H. Kim,

“Using feature selection and classification scheme

for automating phishing email detection,” Stud.

Inform. Contr., vol. 22, no. 1, pp. 61–70, 2013.

[12] O. A. Adewumi and A. A. Akinyelu, “A hybrid

firefly and support vector machine classifier for

phishing email detection,” Kybernetes, vol. 45,

no. 6, pp. 977–994, 2016.

[13] R. Alayham, C. Ren, J. Arshad and A.

Muhammad, "Email Anti-Phishing Detection

Application", Management & Science University,

2019.

[14] M. Husak and J. Cegan, "PhiGARo: Automatic

Phishing Detection and Incident Response

Framework", Masaryk University, Brno, Czech

Republic, 2021.

[15] G. Egozi and R. Verma, "Phishing Email

Detection Using Robust NLP Techniques",

Department of Computer Science University of

Houston, Houston TX, USA, 2021.

[16] Unnithan, Nidhin A., et al. "Machine learning

based phishing e-mail detection." Security-

CEN@ Amrita (2018): 65-69.

[17] L. Zapata, D. Ona, G. Rodriguez, and W. Fuetres,

"Phishing Attacks: Detecting and Preventing

Infected E-mails Using Machine Learning

Methods", Universidad de las Fuerzas Armadas

ESPE, Sangolqui, Ecuador, 2019.

[18] I. Ortiz-Garc, R. Andrade and M. Cazares,

"Detection of Phishing Attacks with Machine

Learning Techniques in Cognitive Security

Architecture", Politecnica Salesiana,Quito, 2021.

[19] Unnithan, Nidhin A., et al. "Machine learning

based phishing e-mail detection." Security-

CEN@ Amrita (2018): 65-69.

[20] Viktorov, Oleg. "Detecting phishing emails using

machine learning techniques." PhD diss., Middle East

University, 2017.

[21] J. Rastenis, S. Ramanauskaitė, I. Suzdalev, K.

Tunaitytė, J. Janulevičius, and A. Čenys,

“Multilanguage spam/phishing classification by

email body text: Toward automated security

incident investigation,” Electronics (Basel), vol.

10, no. 6, p. 668, 2021

[22] Y. Fang, C. Zhang, C. Huang, L. Liu, and Y.

Yang, “Phishing email detection using improved

RCNN model with multilevel vectors and

attention mechanism,” IEEE Access, vol. 7, pp.

56329– 56340, 2019.

[23] Q. Li, M. Cheng, J. Wang, and B. Sun, “LSTM

based phishing detection for big email data,” IEEE

Trans. Big Data, pp. 1–1, 2020.


9 | P a g e



Authors Biography

Muneera Alweheibi, Rawan Alrasheddi, Fay

Aljohar, and Hawra Aljunibi: are currently pursuing

their Bachelors degree in Cyber Security and Digital

Forensics at the department of Networks and

Communications, College of Computer Science and

Information Technology (CCSIT), Imam

Abdulrahman Bin Faisal University, Dammam.

Mainly, their research interests include email security,

Artificial Intelligence and Machine Learning.

Hussain Alattas is currently working in the

department of Networks and Communications

Department, College of Computer Science and

Information Technology, Imam Abdulrahman Bin

Faisal University (IAU), as a lecturer. Hussain has

completed his BS degree in Computer Science from

IAU and MS degree in Cybersecurity and Artificial

Intelligence from The University of Sheffield.

Ghadeer Alazman is currently working in the

department of Networks and Communications

Department, College of Computer Science and

Information Technology, Imam Abdulrahman Bin

Faisal University (IAU), as a teaching assistant.

Ghadeer has completed his BS degree in the science of

Cyber Security and Digital Forensics from IAU.


1 | P a g e

Published In: Journal Of Information And Knowledge Management

Review 1

Intelligent Techniques for Predicting Stock Market Prices: A 2

Critical Survey 3

Abstract: The stock market is a field that many people are interested in, regardless of their occupa- 4

tional background. Individuals who have adequate knowledge can buy shares in the market and 5

generate additional income. Nowadays, the cost of living has increased. Hence, the number of peo- 6

ple who are investing in the stock market is increasing dramatically. While anyone can participate 7

in the stock market at any time, there is no guarantee that they will profit from this investment. The 8

stock market is a risky way to invest, given that it is unknown whether the value of a specific stock 9

will rise or fall. Making stock market predictions using artificial intelligence techniques is a possible 10

way to help people anticipate stock market trends. The current research study showed that many 11

factors impact changes in the stock market’s value in general and in the Saudi Arabia Stock Ex- 12

change specifically. To the best of our knowledge, most previous research only considered historical 13

data for predicting stock market trends. The present study aimed to enhance the accuracy of the 14

daily closing price for three sectors of the Saudi stock market by considering historical data and 15

sentiment data. Several intelligent algorithms were considered, and their performance indicators 16

were discussed and compared. In general, this research study found that more accurate stock mar- 17

ket prediction models can be produced by employing both historical data and sentiment data. 18

19

Keywords: stock market, predictions, artificial intelligence techniques, historical data, and 20

sentiment data. 21

22

1. Introduction 23

Living expenses and taxes have been increasing in recent years, while salaries have 24

become insufficient to meet future needs. Consequently, people are more likely to start 25

new firms or look for extra sources of income. One of the widely utilised methods to ac- 26

complish that is to invest in the stock market, which can provide additional income. How- 27

ever, it requires knowledge of the stock market to correctly predict future stock prices in 28

order to avoid the potential risks. 29

The financial market is a simple system that enables individuals to buy, own and 30

then sell shares at any time with a straightforward process conducted on virtual plat- 31

forms. Although it can be beneficial to do so, investing in the stock market might result in 32

significant losses, particularly if an individual lacks an understanding of stock prices and 33

future forecasts. Furthermore, various factors, including the companies' activities and per- 34

formance, supply and demand and news reports, have significant impacts on prices. 35

These issues necessitate the development of stock price prediction applications to accu- 36

rately estimate stock market prices. 37

Since the beginning of the 21st century, artificial intelligence (AI) technologies, includ- 38

ing machine learning (ML) and deep learning (DL), have become popular and increas- 39

ingly applied in different domains. These strategies focus on employing statistical algo- 40

rithms and exploiting data to build smart systems that can learn, comprehend and act in 41

ways that are indistinguishable from humans in a particular scenario. Consequently, re- 42

searchers agree that they significantly enhance the capabilities of computation, pattern 43

matching and analysing data to extract useful insights quickly and accurately. In the field 44

of the stock market, ML or DL algorithms can be trained with different kinds of data, 45

including historical data, representing a stock’s behaviour, and sentiment data from social 46

media, in order to predict the future prices. 47


2 | P a g e


In this research study, we reviewed and analysed previously published studies that 48

applied AI-based technologies in the field of stock market prediction. Although various 49

literature reviews examining how intelligent-based systems have been used to predict the 50

stock market prices have been published in recent years, none have been as thorough as 51

this one. In this article, we cover 45 research studies published between 2015 and 2021 in 52

the field of stock market prediction. 53

This critical review study also used a novel taxonomy that, to the best of our 54

knowledge, has never been used in earlier studies. It establishes several criteria against 55

which the articles under review can be evaluated and contrasted, including: 56

• The dataset used; 57

• The ML/DL algorithms applied; 58

• The targeted market: local or global; 59

• The kind of features utilised: sentiment data/ historical data/ or both; 60

The performance results were obtained by applying AI techniques. 61

The findings of this literature review point to promising directions for future research 62

and applications in the field of stock market prediction using intelligent algorithms. Re- 63

searchers will be able to use the comparisons and discussions provided in this article to 64

determine which directions to pursue in their research, such as whether to improve intel- 65

ligent-based algorithms or consider other algorithms, which features should be added or 66

removed when building the training dataset, and which evaluation metrics should be 67

used to evaluate the created intelligent systems. 68

The rest of this article is organised as follows. Section 2 presents a literature review, 69

summarising studies focusing on stock markets and the factors that affect stock prices. 70

Section 3 presents a comparison and analysis of the examined research publications, the 71

stock markets they target using ML and DL approaches and their findings. Section 4 pre- 72

sents the study’s conclusions and recommends future research directions. 73

2. Literature Review 74

This section presents an overview of 45 studies that were conducted to predict the 75

future of stock market prices. We reviewed studies that included the idea of applying AI 76

techniques to the stock market to gain a general understanding of the models that were 77

used, to determine how far research has expanded in this field and to identify the ideas 78

that have not been applied in research on the Saudi stock market. Finally, we provide a 79

brief overview of the suggested ideas that we will follow throughout this project. 80

One of the hottest topics that is being discussed is stocks that are being traded, as 81

they are considered to be an additional source of income and savings. The need to increase 82

the source of income has grown after the rise in the cost of living and the increase in the 83

tax burden, as companies make a general appeal for cost-savings to obtain the funds 84

needed for their investments in the form of shares. In the stock market, the profit and loss 85

ratio is based on the participation rate of each individual. Although the stock market can 86

be beneficial to investors, there is a risk in participating in it, as the profit and loss ratio is 87

not guaranteed due to the stock market’s dependence on many factors, such as historical 88

data of the stock, news data, company performance and future expectations, supply and 89

demand and other factors that cause the need for applications that estimate the price of 90

the stock market shares of companies. From this point of view, we will apply AI and DL 91

techniques to estimate stock market prices. AI and DL refer to systems that simulate hu- 92

man intelligence to perform tasks and can be improved based on the information they 93

collect. AI is used in many applications and fields because it provides value to most jobs, 94

companies and industries. After reviewing 45 research studies and applications that ap- 95

plied AI techniques to estimate stock market prices in the future, we found that it pro- 96

vided the results with specific accuracy. Comparing these studies highlights the gap in 97

the research; thus, it will help us develop a new system using AI techniques to estimate 98

stock market prices and close the existing gap in the applied studies [1][2]. 99

The comparison of the research was based on important factors in the application of 100

the system, as follows: 101


3 | P a g e


• Common dataset used; 102

• Common algorithm used; 103

• Implementation of sentiment data to estimate the market price; 104

• Results obtained by applying AI techniques. 105

The results from the comparison and analysis will help researchers apply new ideas and 106

facilitate a new shift in the field of information systems, enabling the development of al- 107

gorithms that can be used as effective techniques for estimating stock market prices. 108

We found some gaps in our literature review. While conducting our research and 109

looking for similar studies that used the Saudi stock market, we noticed that there was a 110

lack of research that analysed sentiment data. 111

Normally, several evaluation indicators are used to evaluate intelligent models. The 112

ones most commonly used to evaluate intelligent stock market models are: 113

• Precision: also known as positive predictive value; it measures the number of 114

correctly predicted cases that turn out to be positive; 115

• Accuracy: the number of correct predictions divided by the total number of pre- 116

dictions; 117

• Correlation: an indicator of the linear relationships (meaning they change at the 118

same rate); it is a common way of interpreting simple relationships without iden- 119

tifying a cause-effect statement; 120

• Recall: also known as sensitivity; it is measured by examining how many posi- 121

tive outcomes can be predicted correctly; 122

• F1-score: a static statistic that expresses the balance between recall and precision; 123

• Error rate: measures the number of patterns that have been predicted incorrectly 124

by the model; 125

• Sum of squared errors (SSE): a weighted sum of squared errors that does not 126

equal constant variance when using heteroscedastic errors; 127

• Mean absolute error (MAE): an average error between the magnitudes of two 128

observations expressing the same phenomenon; 129

• Mean squared error (MSE): the average squared difference between the esti- 130

mated and actual values; it is a quantifier of the quality of an estimator; 131

• Root-mean-square error (RMSE): a commonly used measurement of the differ- 132

ence between the predicted and observed values (sample or population) pre- 133

dicted by the model; 134

• Mean absolute percentage error (MAPE): measures the accuracy of a forecasting 135

method, which is typically expressed as a ratio; 136

• R-squared (R2): depicts how much of a dependent variable's variance is ex- 137

plained by an independent variable or variables in a regression model. 138

139

140

2.1 Research Depending on Historical Data 141

142

One study [3] used the idea of predicting the stock price to such an extent that it can 143

be sold before its worth decreases or bought before the price increases. This study used 144


4 | P a g e


different artificial neural networks (ANNs) to foresee the stock price, but the productivity 145

of forecasting by ANNs relies on the learning algorithm used to train the ANN. This study 146

compared three algorithms: Bayesian Regularization, Scaled Conjugate Gradient (SCG) 147

and Levenberg-Marquardt (LM). It used data (ticks) from 30 November 2017 to 11 January 148

2018 (barring occasions) of Reliance Private. Each day had around 15,000 data focuses, 149

and the dataset contained around 430,000 data focuses. The data were acquired from 150

Thomson Reuter Eikon database. (This dataset was bought from Thomson Reuter). Every 151

change in the price of a stock from one trade to another has a tick that refers to it. The 152

stock price at the beginning of each 15-minute period was extracted from the tick data, 153

which represents the optional dataset run on similar algorithms. Thus, these three algo- 154

rithms have a tick-data utilisation precision of 99.9%. Moreover, for every LM, SCG, and 155

Bayesian Regularization, the exactness over the 15-minute dataset decreases to 96.2%, 156

97.0%, and 98.9%, respectively, which is significantly poor in comparison to the results 157

acquired using tick data. The neural networks (NNs) used in this study are weak; in fact, 158

many other NNs, such as long-short term memory (LSTM), give better predictions. Fur- 159

thermore, applying sentiment analysis can help achieve an additional edge in relation to 160

stock price expectations. 161

Another study [4] focused on the worst prediction accuracy domain, which is the 162

short-term prediction, using time series data of stock prices. An Alpha Vantage applica- 163

tion programming interface (API) was used to access the time series data of 82 random 164

stocks traded at the New York Stock Exchange. (NYSE) The API provides access to daily, 165

weekly and monthly time series data. Since this study used short-term prediction, daily 166

time series data were chosen, which includes the daily opening price, daily high and low 167

prices, daily closing price and daily volume. The study started with a simplified problem, 168

which was predicting whether the prices would increase or decrease in the subsequent 169

days using the stock prices and volumes from the previous days. For this classification 170

problem, logistic regression (LR), Bayesian Network, Simple Neural Network and Sup- 171

port Vector Machines (SVM) with a Radial Basis Function (RBF) kernel were conducted. 172

When using only past price data and technical indicators, the accuracy was found to be 173

70%, which is not high compared to other studies. 174

A study conducted by [5] used ML algorithms, such as Random Forest (RF), K-Near- 175

est Neighbours (KNN), SVM and LR, to evaluate the performance in the field of stocks. 176

That study evaluated the algorithms by assessing performance metrics, such as accuracy, 177

recall, precision and F-score, with the aim of identifying which algorithm most effectively 178

predicted the future performance of the stock market. The dataset is from Kaggle and 179

represents data from the National Stock Exchange of India. That study found that RF had 180

the highest accuracy rate for prediction and the highest recall rate, LR achieved the highest 181

precision and F-score and KNN was the worst performing algorithm among the four that 182

were studied. Overall, RF was the best algorithm, with an accuracy rate of 80.7%. After 183

obtaining the results of the four algorithms, the pros and cons of each technique were 184

identified. Thus, it is easy to determine the best and effective algorithm for the model. 185

One study [6] aimed to apply the KNN and non-linear relapse approaches to antici- 186

pate stock prices for some major companies listed on the Jordanian Stock Exchange. The 187

Jordan Steel Company (JOST), Irbid District Electricity (IREL), Arab International for Ed- 188

ucation and Investment (AIEI), Arab Financial Investment (AFIN) and the Arab Potash 189

Company (APOT) are all listed on that stock exchange to assist investors, decision-makers 190

and clients in making better investment decisions. The study used a dataset of the stock 191

information from 4 June 2009, to 24 December 2009 for five randomly chosen companies 192

recorded on the Jordan Stock Market. Each of these companies has around 200 records 193

with three ascribes, including low price, closing price and high price. The study computed 194

the total squared mistakes, RMSE, and the normal errors for the five companies and iden- 195

tified the contrasts between the anticipated values and the real values in the sample data. 196

It found that the number of errors was small, which demonstrates that the actual value 197

and predicted value are close. According to the results, there is high precision in using the 198

KNN algorithm for forecasting stock values. Then, non-linear regression was applied. The 199


5 | P a g e


outcome indicates that the use of data mining (DM) methods can help decision-makers at 200

various levels when using KNN to examine the data. This study [6] achieved high results 201

using the KNN algorithm as it has a small error ratio, and this yields high precision in 202

contrast with [5], which it did not achieve a high precision result. 203

A DL-based model to predict stock prices was presented in [7]. That study used the 204

historical records of the National Stock Exchange Fifty (NIFTY 50) which contains 50 in- 205

dexes listed in the National Stock Exchange of India from 29 December 2008 to 28 Decem- 206

ber 2018. A multi-step process was used to forecast the opening values of the stock prices 207

of the 50 records. Moreover, the values were predicted week by week, so when a week is 208

over, the actual values of a week are included in the training model before it starts training 209

again. The results using the convolutional neural network (CNN) algorithm show the 210

forecasting performance with a mean of 348.26 in one week, which is better than the mean 211

of 407.14 for two weeks. 212

The study conducted by [8] predicted the ability of various well-known forecasting 213

models, including dynamic versions of a single-factor Capital Asset Pricing Model 214

(CAPM)-based model and Fama and French's three-factor model, to close the gap in the 215

literature. The dataset was collected from the Shanghai Stock Market and it compared the 216

predicting performance of each of the six models with the performance of an ANN model 217

using the same predictor variables; however, it relaxed the model linearity assumption. 218

Surprisingly, there were no statistically significant differences between the CAPM and the 219

three-factor model in terms of forecasting accuracy. Furthermore, each ANN model out- 220

performed the equivalent linear model, showing that NNs might be a valuable tool for 221

predicting stock prices in emerging markets. On average, the overall accuracy of the pro- 222

posed method is equal to 0.0113 MAD, 0.3118 MAPE and 0.2807 MSE. As discussed in [3], 223

by utilising an ANN algorithm, the model provides more accurate results. 224

Another model that analysed the stock market and identified nonlinear relationships 225

between the input data and the output data was proposed in [9]. Two types of ANN algo- 226

rithms were used in this study, RBF and Multi-Layer Feed Forward (MLFF using the 227

Shanghai Stock Exchange composite index in China. The reason for using two ANN tech- 228

niques is that an RBF network can deal with nonlinear functions and operate with the 229

complexity of analysing the rules and laws in the system, while MLFF is used to deal with 230

the complex nonlinear relationship between the input and output data. This study found 231

that RBF outperformed MLFF because the RBF's error is substantially smaller. The appli- 232

cation provided an excellent comparison of two types of ANN algorithms. 233

The study conducted by [10] focused on the efficacy of DL in predicting one-month- 234

ahead stock returns in a cross-section of the Japanese Stock Market. NNs have been used 235

in several studies on stock return predictability. NNs have also been used to make indi- 236

vidual stock return estimates. The MSCI Japan Index dataset consists of data from Decem- 237

ber 1990 to November 2016 and contains 319 indexes. The study used ANNs, Support 238

Vector Regression (SVR) and RF. The result shows that deep NNs outperform shallow 239

NNs, in general, and the top networks also beat typical ML models. Indeed, the findings 240

suggest that DL shows potential as a sophisticated ML method for predicting cross-sec- 241

tional stock returns. A future study could include the use of RNN, which is designed to 242

handle time series data. An analysis of several DL models is also predicted to improve the 243

accuracy of stock return prediction in the cross-section data. 244

Another study [11] aimed to construct a novel ensemble ML framework for daily 245

stock pattern prediction by combining traditional candlestick charting with the latest AI 246

methods. The Chinese Stock Market dataset was used in this research, with a total of 247

65,000 rows of data in each round. A total of six ML models were used in this study: LLR, 248

SVM, KNN, RF, Gradient Boosting Decision Tree (GBDT) and LSTM. After comparing the 249

results of each of these models, RF and GBDT showed a good predictive ability for short- 250

term prediction, whereas the LR prediction level needs to be improved and KNN and 251

SVM only fit in some patterns. The LSTM model has more advantages as a DL, but those 252

advantages were not fully discussed. Overall, the model had an accuracy greater than 253

52%, and an F1-score greater than 50%. This research provides useful information and it 254


6 | P a g e


is distinct from other studies conducted on the basis of stock market prediction, because 255

it shows in detail the results of each model that used different algorithms. 256

A study conducted by [12] applied an ML algorithm and time series forecasting using 257

Microsoft Excel as the best statistical tools for graphic and tabular representation of the 258

prediction results. That study used data from Yahoo Finance for Amazon (AMZN) stock, 259

AAPL stock and Google stock datasets. They focused on using LR, three-month moving 260

average (3MMA) and Exponential Smoothing (ES) algorithms. Three different prediction 261

methods were considered. Of them, ES based on LR showed the best results with a 16.62 262

average absolute error. The study is distinct from others as it also applied a time-series 263

analysis to predict the stock market prices for the next month. 264

In [13], two effective models were developed using ANN and SVM classification 265

techniques to predict the direction of stock price index movement. Then, the ability to 266

anticipate the direction of movement in the daily Istanbul Stock Exchange (ISE) National 267

Index was tested. The data were from the ISE National Index's daily closing price move- 268

ment from 2 January 1997 through 31 December 2007. The average performance was 269

found to be better for the ANN model (75.74%) than the SVM model (71.52%). The pre- 270

diction performance of these models can be improved in two ways. The first way adjusts 271

the model parameters by conducting a more sensitive and thorough parameter setting. 272

The second way is to employ additional macroeconomic variables, such as foreign ex- 273

change rates, interest rates and the consumer price index, as inputs to the models. 274

One study [14], proposed a model to understand the financial market and build a 275

neural model for the financial market theory with respect to technical analysis, fundamen- 276

tal analysis and time-series analysis. That study used the feedforward multilayer percep- 277

tron ANN algorithm. This algorithm is used because of its efficiency in predicting a time 278

series and its ability to learn and recognise non-linear data. The authors conducted a sur- 279

vey to gather input from qualified professionals on the models, techniques and indicators 280

used in the pricing of stocks. A questionnaire was sent to 50 investors and analysts work- 281

ing in the stock market. The datasets were obtained from Economatica, Brazil’s Central 282

Bank, the São Paulo Stock Exchange and Thomson Reuters. The result is based on the set 283

of error metrics with a window size equal to 3, as it presents a Prediction of Change in 284

Direction (POCID) correct rate of 93.62% and a MAPE of 5.45%. The results could be fur- 285

ther improved by expanding the algorithms used to obtain high accuracy and discover 286

the best algorithm for forecasting stock market prices and trends. 287

Another study examined the prediction power of NN modelling and SVM to forecast 288

Russian stock prices [15]. The dataset was consisted of the daily Moscow Interbank Cur- 289

rency Exchange (MICEX) stock price index, as well as some technical and fundamental 290

indicators from 2002 to 2016, based on statistical and analytical methods. Datasets are used 291

for training, testing and verification in Python for ML. Feedforward NNs were used to 292

predict the MICEX index. Moreover, a back propagation (BP) algorithm was used to train 293

it. The study used the activation function as its baseline and the dependent and independ- 294

ent variables were normalised to the interval [-1,1]. To decrease the potential for overfit- 295

ting problems, the data were split into a 60% training sample and a 20% testing sample. 296

To determine which parameters of the learning algorithm and NN architecture for each 297

sample are optimal, training samples and testing samples were used. The performance of 298

the NN was also evaluated using a validation sample (20%). The NN’s optimal learning 299

parameters were found empirically using a grid search. An optimal configuration was 300

also found by training and testing processes in SVM. Data normalisation in SVM was ac- 301

complished using transformations. The prediction performance of the NNs and SVM was 302

compared based on MSE, RMSE, MAE, MAPE, R2 and the calculated coefficient of deter- 303

mination (cR2). B SVM was found to have a higher predictive power than NN modelling. 304

305

306

307

A genetic algorithm (GA) proposed in [16] was used to forecast prices and trends for 308

the India Stock Market. The dataset was extracted from the India TCS Stock Market for 309


7 | P a g e


trading values for 259 days, including the opening, closing, lowest and highest prices of 310

each day's trading. The GA is a search algorithm that combines the mechanics of selection 311

and genetics. There are three types of genetic operations: crossover, mutation and selec- 312

tion. Historical data were used to predict future search points with improved performance 313

efficiently. The study was conducted using a time-series analysis, which identifies pat- 314

terns in statistical information by returning information at regular intervals of time. To 315

make a prediction for the opening price of the next week, the closing price values of the 316

previous week are taken by ignoring the negative ‘ve’ sign, since the opening price is al- 317

ways higher than the closing price. In contrast, when predicting the closing price from the 318

opening price the sign changes from a positive ‘ve’ to a negative ‘ve’. However, the sign 319

does not change when predicting low and high prices. The Chi-square test was used to 320

determine whether the prediction was significant or merely a coincidence. According to 321

the test findings, time series and GA significantly improved the prediction system's accu- 322

racy by 99.87%. 323

Another study [17] proposed a practical method for predicting stock development. It 324

did not present any numerical result as the aim was to present the most appropriate anal- 325

ysis for anticipating the stock market. A dataset from the previous year's stock market was 326

employed and divided into training and testing data to improve the accuracy. The RF 327

algorithm and the SVM algorithm were both considered. The results of the calculation 328

showed that the RF algorithm performed better in predicting a stock's market price, which 329

was proven in [5] and [11], as both studies achieved higher performance using RF. 330

Since stock price prediction using time series forecasting is one of the most complex 331

challenges in the financial field, a method was developed in [18] for predicting stock price 332

and time series using a hybrid method of GA and ANN techniques. That study used data 333

from Apple, Pepsi, IBM, McDonald's and LG. Compared to traditional models, the pro- 334

posed solution exhibited a 99.99% improvement in SSE and a 99.66% improvement in time 335

when a hybrid model of GA and BP was applied to a dataset of Apple stocks. When the 336

Pepsi dataset was used, the approach had an SSE of 0.0121281374; traditional methods 337

without using GA had an SSE of 0.4790571631. That study yielded an SSE accuracy of 338

99.42% and a time reduction of 88.75%. Their method could be further improved by com- 339

bining it with other methods, such as SVM or decision tree (DT), or by expanding their 340

study to include a time-series analysis as was done in [16]. 341

In [19], it was reported that ANN is suitable for stock market prediction since it is a 342

popular way to identify unidentified and unseen patterns in data. That study was divided 343

into two modules; one module was for training and the other was for predicting the stock 344

price based on the previous training. A method was proposed to predict the share price 345

using a BP algorithm and an MLFF network. The dataset was obtained from ACI Pharma- 346

ceuticals. Using two input datasets and five input datasets, a difference was found be- 347

tween the anticipated and actual stock price. The error percentage between the predicted 348

price and the actual price decreased when the model had more training. When using five 349

input datasets, the highest error rate was 3.28% and the lowest was 0.12%. The method 350

used to achieve a more error free prediction was done by training the system with more 351

input datasets. Another method could also be conducted to yield better results; for exam- 352

ple, in [18], ANN was also used, but it was combined with GA to improve the results. 353

Another study [20] proposed an intelligent stock market forecasting system using the 354

ability of ANN and a fuzzy inference system. The goal was to notice the patterns in non- 355

linear and disordered systems. The dataset was from BEXIMCO Ltd. Using a model that 356

combined an NN and fuzzy logic, a total of nine inputs were used from the prediction 357

dataset to compare the actual price and the predicted price. The highest error rate of this 358

model was 4.8895% and the lowest was 0.3734%. Training the model with more data, as 359

was done in [19], could improve its performance and generate a more error free predic- 360

tion. 361

The study conducted in [21] utilized SVM to forecast ISE prices in Turkey. The study 362

proposed a method for learning to predict stock price returns by considering a binary 363

classification problem (positive and negative). Positive return forecasts were represented 364


8 | P a g e


by a class label of +1 and negative predictions were represented by a class label of -1. 365

Learning the learned model involves the use of weighted SVM, RF classifi-ers, Relevance 366

Vector Machine (RVM) and Multiple Layer Perceptron (MPL). A three-layer feed forward 367

technique was implemented with 10 neurons as an input, a neural layer for each technical 368

parameter, and a neuron as an output layer to show the predicted result. To update the 369

weights, the tangent sigmoid was used as a transfer function in conjunction with a gradi- 370

ent descent algorithm. The error and output of the initial network were calculated during 371

adaptive gradient descent. In this way, a near-optimal learning rate for the local environ- 372

ment can be obtained. Additionally, a higher learning rate is guaranteed if stabilised learn- 373

ing occurs. The accuracy of the proposed method was reported to be 70%. In contrast to 374

other studies that utilised the SVM algorithm and produced high accuracy, this study 375

should improve its methods to achieve higher accuracy. 376

Another study [22] proposed a method that can forecast stocks from different mar- 377

kets and industries and predict the trend using ML algorithms, such as polynomial re- 378

gression and LR, in addition to learning techniques for predicting a time series using two 379

special types of NN recursions: spoken short-term memory and LSTM. The historical in- 380

formation contains data for the daily low, high, closing and opening prices and the vol- 381

ume of each stock. The dataset consisted of the five-year window of Alibaba, VinGroup, 382

Reliance and PepsiCo to guarantee that both bullish and bearish trends in this period 383

would be investigated. First, LR and polynomial regression were used to complete the 384

regression analysis and predictive analysis of the stock information. Second, the LSTM 385

model was used according to the qualities of stock market data because of its excellent 386

performance in successive data processing, choosing the Stochastic Gradient Descent 387

(SGD) and Adaptive Moment Estimation (Adam) as the optimizers. Finally, this study 388

used the LSTM combined model enhanced by a one-dimensional CNN (CONV1D) for 389

forecasting, which works on the exactness of the expectation model because there is a high 390

error rate in LSTM. The test results confirmed the efficacy of the original LSTM network 391

by adding two CONV1D layers, which helped improve the overall accuracy. Moreover, 392

the RMSE and MAPE values were smaller when using the Adam enhancer than the SGD 393

enhancer. Thus, the CONV1D-LSTM model improved by Adam is more reasonable and 394

produces better prediction with an accuracy of 54.17% based on the Alibaba dataset, 395

51.56% based on the PepsiCo dataset, 51.38% based on the VinGroup dataset and 50.01% 396

based on the Reliance dataset. This study achieved high accuracy by improving the origi- 397

nal LSTM to create a CONV1D-LSTM model in contrast to [11], which only used the orig- 398

inal LSTM without any adjustments. 399

One study aimed to predict future values of portfolios using an ML algorithm de- 400

pendent on LSTM and RNN to estimate the changes in the closing prices for a portfolio of 401

resources [23]. The objective was to obtain an accurate, trained algorithm. The study used 402

the datasets of two stocks at New York Stock Exchange (NYSE) consisting of the daily 403

opening prices. Two stocks (Google and NKE) are extracted from Yahoo Finance. The data 404

for the Google series covers the period from 19/8/2004 to 19/12/2019 and the data for NKE 405

covers the period between 4/1/2010 to 19/12/2019. LSTM and RNN were applied to build 406

the model, which used 80% of the data for training and 20% of the data for testing. The 407

test results were strongly influenced by both the number of epochs and the length of the 408

data. For the training data, 12 epochs, 25 epochs, 50 epochs and 100 epochs were used. 409

That study found that training with fewer data and more epochs improved the testing 410

results and also improved the forecasting and prediction values, depending on the da- 411

taset. Thus, the model can trace the evolution of the rates of opening prices for both assets. 412

In the future, the study will work to identify the mix of session data length and the number 413

of training epochs that best suit their resources and augment the accuracy expectations. 414

Another study [24] worked to develop a model dependent on technical indicators 415

with LSTM to forecast the price of a stock at 1 minute, 5 minutes and 10 minutes. High- 416

frequency data were used by combining LSTM and classical financial models to predict 417

the closing price. The dataset from Kaggle consisted of S&P 500 intraday trading data. The 418

original data files contained 484 observations. One observation has a time stamp, as well 419


9 | P a g e


as opening, low, high and closing prices and volume. Data from 11/9/2017 to 16/2/2018 420

were used and had a total of 43,148 sequence data. The dataset was divided into a training 421

set and a validation set. The period of the training set was 11/9/2017 9:30 A.M. to 17/1/2018 422

11:50 A.M. The period of the validation set was 17/1/2018 11:51 A.M. to 16/2/2018 03:59 423

A.M. Following that, experiments were conducted to predict the price at 1 minute, 5 424

minutes and 10 minutes. The basic idea was to check how close each model is to reality to 425

know the extent of the risk that the user may be exposed to while predicting the stock 426

price x-minutes before. For each observation, there is a model with and without technical 427

pointers to improve the analysis. Therefore, this study coincides with establishing the in- 428

fluence of technical indicators in forecasting because the accuracy stays below 50%. The 429

model also affirms that the closing price can be predicted 10 minutes before closing and 5 430

minutes before closing, with the best performance seen 1 minute before closing, without 431

the use of technical indicators. This study may need to focus on sampling and back-testing 432

to best dominate this domain. 433

The capacity of ANN to forecast the everyday NASDAQ stock exchange rate was 434

examined in [25]. That study used short-term historical stock prices and the day of the 435

week as inputs. Using NASDAQ data from 28 January 2015 to 18 June 2015, they applied 436

daily stock exchange rates to develop a powerful model. The initial 70 days (28 January to 437

7 March) were chosen as the training datasets and the last 29 days were used for testing 438

the model’s prediction ability. Networks for the NASDAQ index that forecast two kinds 439

of input datasets (4 days earlier and 9 days earlier) were developed and approved. The 440

determination coefficient (R2) was used to evaluate the performance of the ANNs and the 441

MSE of the modelled output. The study applied the OSS training technique and TANGSIG 442

transfer function in a network with 20-40-20 neurons in hidden layers. The result was a 443

streamlined prepared network with R2 values of 0.9408 for the approval dataset. In this 444

dataset, most of the R2 values for the networks with the OSS training method and TANG- 445

SIG transfer function could be obtained when the number of neurons was 40-40 and the 446

number of hidden layers was 2. For 9 earlier working days, a network with 20-40-20 neu- 447

rons in the hidden layers OSS training method and the LOGSIG transfer function, the up- 448

graded network achieved an R2 of 0.9622. The results show that there is no difference 449

between the prediction ability of the 4 and 9 prior working days as the input parameters. 450

While ANN was used in [18], it was combined with GA to achieve high results. 451

452

The study conducted by [26] proposed trading strategies by combining technical 453

analysis indicators and data on stock market returns with ML approaches. To test the rec- 454

ommended algorithms RF, LR, ANN and SVM were chosen. The data set from Guaranty 455

Trust Bank contained stock data from the Nigerian Stock Exchange (NSE). Therefore, the 456

stacking technique was implemented to discover which of the four algorithms, when used 457

as the top layer and the remaining as the second layer, could effectively predict the stock 458

returns values. When compared to the real value of stock returns, the results of the exper- 459

iment showed that the top layer of the RF algorithm can forecast buy and sell signals. The 460

authors may expand their research by learning from social media news, which has a high 461

correlation with stock prices and market status. 462

The model proposed in [27] was used to predict Saudi stock market prices by em- 463

ploying AI and soft computing techniques. The proposed model is based on Saudi Stock 464

Exchange historical data and it was tested on three companies: SABIC, Saudi Telecommu- 465

nication Company (STC) and Al-Rajhi Bank. The proposed ANN model predicts the next 466

day closing price stock market value with a very low RMSE of 1.8174, a Mean Absolute 467

Deviation (MAD) of 18.2835, a MAPE of 1.6476 and a very high correlation coefficient of 468

up to 99.9%. The ANN model was used because of its ability to learn, memorise and create 469

relationships among data. It has been proven to be a good tool for predicting the Saudi 470

stock market prices, in a context in which little recent research has been conducted. This 471

study could be further strengthened by applying more than one algorithm and comparing 472

the results to achieve the highest possible accuracy rate. 473


10 | P a g e


Another study [28] used the BP algorithm to predict the Saudi stock market prices by 474

applying different technical indicators, training and transfer functions to evaluate the per- 475

formance of NN methods. Real historical data from the Saudi stock market and OPEC 476

crude oil prices from TADAWUL Stock Market Exchange and historical oil price data 477

from 2003 to 2015 were used to evaluate the effectiveness of the NN methods. The best 478

training function was found to be MATLAB trainbr; it had the highest accuracy in com- 479

parison to other training functions with the ANN algorithm. The highest accuracy of the 480

ANN model using trainbr was 75.7%. 481

One study [29] aimed to test the Saudi stock market weakness using the Kolmogo- 482

rov–Smirnov test and the Market Weak Form Test results for an active market hypothesis. 483

The RNN proposed to create a trading signal with a long momentary memory architecture 484

to predict the following day trading of a few shares on the Saudi Stock Exchange. The 485

Saudi stock lists were then utilised in sequence with a trading algorithm to purchase and 486

sell shares depending on three elements: the current number of shares owned, the current 487

available balance and the current share value. The dataset for the Saudi Stock Exchange 488

consists of three stock shares. An example of chronicled values with 55% exactness is data 489

from June 2018 to August 2019 for SABIC, Alinma Bank and Al-Rajhi Bank. The 55% ac- 490

curacy result and an investment gain of 23% was satisfactory , in comparison to the results 491

obtained with the buy-and-hold trading method, which achieved a 1.2% investment gain. 492

However, more factors could be considered in this study, such as the Fibonacci retrace- 493

ment and implementing a component choice strategy to select the best element among the 494

introduced features. Or, the study could consider trading strategies to prepare an NN and 495

improve a trading agent as opposed to depending on the forecast of future returns. 496

Another study conducted in Saudi Arabia [30] aimed to enhance the forecasting ac- 497

curacy of the Saudi Arabia Stock Exchange (TADAWUL ) data patterns. The study used 498

datasets from TADAWUL, the Saudi Authority for Statistics and the Saudi Central Bank. 499

With a total of 2026 records, the MODWT functions (a mathematical model based on five 500

functions) was combined with the adaptive network-based fuzzy inference system (AN- 501

FIS) model to develop with this model. The results of this model are more accurate (99.1%) 502

than those of traditional models. This forecasting model is capable of decomposing in the 503

stock markets; unlike traditional models, this model excelled because it resulted in a high 504

accuracy percentage as it was a combination of MODWT and ANFIS. 505

A model was proposed in [31] to predict the Saudi stock price trends with regards to 506

its earlier price history by combining a discrete wavelet transform (DWT) and RNN. Past 507

data from TADAWUL was used, with a total of 130,000 records. Two models were pre- 508

sented in this study, DWT+RNN and Auto Regressive Integrated Moving Average 509

(ARIMA). The purpose of using ARIMA was to compare the proposed method 510

(DWT+RNN) with a traditional prediction algorithm (ARIMA). The MAE for the pro- 511

posed method was 0.15996, MSE was 0.03701, and RMSE was 0.19237 RMSE; this demon- 512

strates a significant improvement in comparison to ARIMA, which had an MAE of 6.60949 513

MAE, an MSE of 76.5758 and an RMSE of 8.75076. This integration method could be used 514

to formulate better and improved techniques to reduce the risks of investing and assist 515

investors in making stock-buying and selling decisions. The prediction of this model could 516

be further increased by considering other factors that might affect the accuracy of the 517

price, which is also the case for other research studies. 518

Another method to predict the Saudi stock market was proposed in [32], which con- 519

sidered people’s sentiments about their financial decisions. This study used TADAWUL l 520

All Share Index (TASI) and Global Data on Events, Location and Tone (GDLET) Google 521

database (collection of news from all over the world from different types of media, includ- 522

ing TV, podcasts, radio, newspapers and websites). The goal was to use a time-series anal- 523

ysis to predict the Saudi Stock Market Index by incorporating the GDELT dataset with the 524

TASI. Statistical and ML approaches were used. Of all the models that were tested in this 525

study, LSTM (a special kind of RNN with the ability to learn long-term relationships) had 526

the best performance with an MAE of 0.59. LSTM can give very accurate forecasts, as it 527

had a very low MAE. The study also mentioned how challenging it is to forecast stock 528


11 | P a g e


market prices since a stock market is dependent on multiple factors, in addition to histor- 529

ical data, that vary in type and extraction complexity. Financial markets can be influenced 530

by economic factors and non-economic factors. The research done in this study was com- 531

plete as more than one model was tested before concluding which model was best; how- 532

ever, it might have been lacking in other areas as a TADAWUL dataset was also used in 533

[31] where it achieved a higher MAE. 534

The study conducted by [33] used several ML algorithms to predict stocks in the 535

Saudi Arabia Stock Exchange, such as Multilayer Perceptron, SVM, Ada-boost, Naive 536

Bayes, Bayesian Networks, KNN and RBF, to forecast the Karachi Stock Exchange price 537

and test the algorithm on Saudi stocks for TASI. The dataset from 10 companies was ob- 538

tained from the Karachi Stock Exchange over a period of six months, from April 2013 to 539

September 2013. Several matrices were used to compare the performance of these algo- 540

rithms, including MAE, RMSE and accuracy. The analysis was categorised by time, most 541

recent data used for testing and historical data for training. The training dataset was split 542

into an 80% training set and a 20% testing set. A 2-fold cross-validation method was per- 543

formed over 80% of the training to determine the optimal parameters. It was found that 544

Ada-boost, Multilayer Perceptron and Bayesian Network were more accurate than the 545

other tested algorithms. 546

547

2.2 Research Depending on Sentiment Data 548

549

To discover the effectiveness of social media of Hewlett-Packard Corporation and 550

news data on stock market prediction, [34] used this data with ML algorithms, such as 551

SVM, and Naive Bayes. The algorithms were applied on 10 subsequent days and S&P 500 552

index price data were gathered for 2 years from 1 July 2016 to 30 June 2018. To improve 553

the performance and quality of stock predictions, spam tweets were scaling down and 554

duplicate tweets were removed from the social media datasets. Moreover, the ML model 555

can provide raw text data in the form of tweets and news headlines. These data are not 556

understandable by ML algorithms and need to be pre-processed. NLP is used to analyse 557

social media and news data to find (positive, neutral or negative) sentiments. Then, ML 558

algorithms are used to learn the relationship between the sentiments of the text and the 559

stock market trends. The result show that the highest prediction accuracies of 80.53% and 560

75.16% are achieved using social media and news, respectively. 561

Since few studies analysed sentiment Arabic tweets on the stock market, a model was 562

proposed to analyse the relationship between Saudi Twitter posts that contain sentiment 563

data with people’s opinions and the Saudi stock market [35]. The data were collected us- 564

ing a desktop application written in C# (Twitter data grabber) to save and label the tweets 565

as positive, negative or neutral, and discard irrelevant data. The tweets were obtained 566

from the Mubashir website in Saudi Arabia. The dataset contains 3335 records collected 567

from 17/3/2015 to 10/5/2015 of all share sectors of the Saudi Arabia Stock Exchange that 568

exist on the TADAWUL website. After labelling the data, normalisation occurs as follows: 569

positive tweets (1), negative tweets (-1) and neutral tweets (0). Moreover, three algorithms 570

were used to build the model: Naive Bayes, SVM and KNN. The evaluation was done 571

using precision and recall methods. Term Frequency-Inverse Document Frequency (TF- 572

IDF) as used to extract the features. A clear correlation was found between news pub- 573

lished on social media (especially Twitter) and the Saudi stock market. This proved the 574

high impact of sentiment data on the Saudi stock market. 575

576

2.3 Research Depending on Both Historical and Sentiment Data 577

578

The system proposed in [36] integrates ML, mathematical operations and some other 579

factors, including news sentiment, to gain better forecasting accuracy and generate prof- 580

itable trades. Two sources of information were used in this study: market news sentiment 581


12 | P a g e


and historical prices. The study assumed that market news can be classified as positive, 582

negative and neutral. The news sentiment dataset includes information about the number 583

of positive and negative words in news articles related to the companies considered in the 584

training dataset. That dataset was collected from the Reuters platform; it consists of data 585

related to several major companies, such as Apple (AAPL), Google (GOOGL), Amazon 586

(AMZN) and Facebook (FB). Several approaches were used to build the intelligent mod- 587

els: RNN, Deep Neural Network (DNN), SVM and SVR. Various evaluation metrices were 588

used to evaluate the created models. The most important evaluation metrices considered 589

in this study are the directional accuracy (which analyses the direction of the predicted 590

value with respect to yesterday’s closing price), precision (which measures the relevancy 591

of the result), recall (which measures how many true relevant results returned) and F- 592

score (which measures the weighted average of precision and recall). All the intelligent 593

approaches showed high forecasting accuracy. However, SVM showed the best accuracy 594

(82.91%). 595

To make the process of investing in the stock market less time-consuming, simple, 596

easy and less stressful, a model to predict stock fluctuations was proposed [37]. This 597

model will help new investors obtain a deeper understanding of the stock market. The 598

data were obtained from Yahoo Finance. The model was created by considering several 599

features, including stock market prices for the previous year, past values of the currency 600

and commodity markets, historical news headlines, sentiment data and international 601

stock market data. All the smart algorithms that were considered in this study—ANN, 602

RF, SVM, KNN and LSTM—demonstrated high accuracy in predicting stock fluctuations; 603

however, RF had the best accuracy: 80%. 604

Prediction of stock prices with AI and social media was conducted in [38]. The main 605

goal was to create an NN based on LSTM that can forecast stock market movements based 606

on user tweets. Additionally, that study worked on developing an RNN, an LSTM varia- 607

tion capable of predicting short-term price fluctuations. The popularity of RNNs in NLP 608

and stock prediction tasks is attributed to the fact that they consider the temporal effect of 609

events, which is a significant advantage over other NNs. DL methods were employed for 610

this task since hidden layers may take advantage of the inherent relational complexity and 611

extract these implicit links. Consequently, the LSTM structure was chosen as the primary 612

model. The stock price data were the same as the data used in [37] and the Twitter data 613

was obtained from Follow the Hashtag. The degree of association between the sentiments 614

conveyed via tweets and the direction of the stock prices were also explored with the use 615

of a popular sentiment analysis tool known as VADER. This was done to compare the 616

results with those from the LSTM architecture. It was found that VADER was unable to 617

extract any strong relationships between social sentiment and market direction. The mod- 618

el's final testing accuracy was 76.14%. Although the level of accuracy is excellent on its 619

own, Twitter datasets from other technical businesses must be reviewed and compared to 620

the findings of this study. This relative comparison will enable a more realistic assessment 621

of the LSTM network's performance in a broader context. 622

Twitter attribute information was used to predict stock prices in [39]. An NN-based 623

model with several layers of LSTM was utilised. The model was trained using Twitter 624

attributes as well as historical stock values. Twitter API was used to collect data and to 625

retrieve the needed attributes. The results showed that adding Twitter attributes to the 626

model yields a 3% improvement in MSE, which was about 0.002 MSE. Consequently, an- 627

other experiment was done that used the sigmoid function on the follower count; it re- 628

sulted in an MSE of 0.14. Further improvements were made by using a sigmoid function 629

when scaling the tweet follower count attribute, which yielded an MSE of 0.13. This study 630

took Twitter information into consideration, and it succeeded in showing the degree to 631

which the accuracy rates differ when using Twitter attributes with the historical stock rec- 632

ords in the model. 633

A simple LSTM model with complex level embeddings for securities market forecast- 634

ing, while utilising monetary news as predictors, was applied in [40]. First, an RNN model 635


13 | P a g e


was used for the forecast. Then, a neural language model was used to build a representa- 636

tion for the information. The model was assessed on a dataset of monetary news from 637

October 2006 to November 2013, which was made accessible by Ding et al. (2014) and 638

gathered from Reuters and Bloomberg. The stock price information for all S&P 500 com- 639

panies and the S&P 500 index was acquired from Thomson Reuters Tick History. The test 640

results of the model S&P 500 Index expectation, in relation to the precision of foreseeing 641

stock price development on the test dataset, showed an accuracy of 63.34%. The test con- 642

sequences of individual company forecasts had an accuracy of 64.74%. The character-level 643

language model pre-training performs just like all of the different models, but with the 644

benefit of being easier to implement because it does not have a module for modelling 645

events. Based on the outcome, the study may need to test the utilisation of the character 646

embeddings with more difficult designs and possibly the expansion of different infor- 647

mation to make richer feature sets. 648

A hybrid approach for stock price movement prediction using ML and DL was pro- 649

posed in [41] using the daily historical data of the NIFTY 50 index of India from 2/1/2015 650

to 28/6/2019. Based on the data collected, this study built various predictive models using 651

ML with SVM and ANN algorithms. Moreover, in this research, Twitter data were used 652

to gather public sentiment about stock prices and compare it with the market sentiment. To 653

predict the price movement patterns, several classification techniques were used. In the 654

classification approach, “0” indicated the negative value while “1” indicated the positive 655

value. Hence, if the forecast model expects an increase in the value on the next day, the 656

value of the next day would be “1”. A predicted negative value on the next day would be 657

indicated by a “0”. Additionally, an LSTM was built using a DL network for predicting 658

the stock price, and the accuracies of the ML models and the LSTM model were compared 659

to find the most accurate approach. The result shows a correlation of 0.99, an MAPE of 660

10.75 and Matched Cases of 80%. In comparison to other studies that used the same da- 661

taset, this research reported great performance measures. 662

663

In [42], stock price movements in Indonesia were predicted based on sentiment anal- 664

ysis, technical analysis and fundamental analysis using SVM and NLP. That study had 665

two objectives: predicting the movement of stocks in Indonesia based on sentiment anal- 666

ysis, technical analysis and fundamental analysis using SVM, and measuring their impact 667

on the prediction. The dataset was obtained from nine companies: Astra Agro Lestari 668

(AALI). Astra International (ASII)), Bank Central Asia (BBCA), Merdeka Copper Gold 669

(MDKA), Pakuwon Jati (PWON), Telekomunikasi Indonesia (TLKM), Chandra Asri Pet- 670

rochemical (TPIA), United Tractors (UNTR) and Unilever Indonesia (UNVR). The news 671

sentiments were collected from online media outlets, such as the CNBC Indonesia Twitter 672

account. The historical data of stocks were obtained from the Yahoo Finance website. NLP 673

was used to process human language in sentiment analysis while SVM was used to build 674

the prediction model since it is not readily impacted by data outliers and it can understand 675

complex stock price movement data patterns. The average accuracy rate obtained was 676

65.33%. However, this study could have applied more algorithms to obtain the highest 677

possible accuracy rate. 678

In [43], DL was used to predict the stock market price. That study examined various 679

tactics for projecting future stock prices and used a pre-built model that is tailored to the 680

Moroccan Stock Market as an example. It also compared the outcomes of a single LSTM 681

model, a stacked LSTM model adapted to the Moroccan Stock Market using the BMCE 682

BANK stock price data set and a hybrid model that uses both stacked auto-encoders and 683

sentiment analysis. It was found that DL significantly aided in solving the problem of 684

stock market forecasts; therefore, the study employed LSTM and NNs. The results showed 685

that combining LSTM networks with stacked auto-encoders and sentiment analysis pro- 686

duced results that are suitable for live trading. However, the difficulty was with the 687

amount of data available, which was insufficient for an optimal and profitable DL model. 688

This limitation may be overcome by using data augmentation techniques on current data 689

sets to increase their size and make them suitable for DL projects. 690


14 | P a g e


ANNs were implemented in [44] to examine the relationship between public senti- 691

ments and the predictability of future stock price development. Three datasets were used 692

from January 2010 to September 2019. The first dataset was the authentic stock prices 693

(Dstock) of three companies (GCB, MTNGH and TOTAL) recorded on the Ghana Stock 694

Exchange. The second was a disorderly dataset, which included 2184 tweets, web news 695

(myjoyonline.com, ghanaweb.com and graphic.com.gh) and 1581 posts. The third dataset 696

contained 263 records from Google Trends, provided by Google. They anticipated the fu- 697

ture stock incentive for a period of 1 day, 7 days, 30 days, 60 days and 90 days. The result 698

was an accuracy ranging from 49.4% to 52.95% based on Google Trends, 55.5% to 60.05% 699

based on Twitter, 41.52% to 41.77% based on a forum post, 50.43% to 55.81% based on web 700

news and 70.66% to 77.12% based on a combined dataset. This study did not achieve as 701

good accuracy as that reported in [3], which also used the ANN algorithm. 702

Another study used the fundamental analysis technique to find future stock trends 703

by considering news articles about Apple published from 1 February 2013 to 2 April 2016 704

[45]. The company’s data is prime and it tries to classify news as good (positive) and awful 705

(negative). The study implemented three classification models and tested them under var- 706

ious scenarios. The findings show that RF returned the best results for all the experiments, 707

with an accuracy ranging from 88% to 92%. The SVM followed with an accuracy of 86%. 708

Naïve Bayes had an accuracy of 83%. The model was able to effectively predict the stock 709

trend in any news article. This presumes that stock patterns can be anticipated by utilising 710

news articles and previous price history due to the model’s high accuracy. However, this 711

study may need to broaden its research focus by adding more company data and by look- 712

ing at the forecast precision. For companies where financial news is inaccessible, the study 713

may use Twitter data for a comparable accuracy, or it can incorporate similar methodolo- 714

gies for algorithmic trading. 715

A model that integrates social media sentiments to predict stock price movements for 716

rising and falling stocks was developed in [46]. SVM handles data efficiently in high di- 717

mensionality and has been shown to perform well in classification. Thus, this study chose 718

SVM with a linear aspect as the prediction model. Several datasets were used for the 719

model. The first was a dataset of historical data prices. The second dataset was the mood 720

information data from social networking sites. Six sets of features were designed to assess 721

the effectiveness of sentiment analysis on message boards: human sentiment, aspect- 722

based sentiment, price only, sentiment classification, joint sentiment/topic-based method 723

and linear discriminant analysis. The results obtained using a topic-sentiment feature 724

were slightly better than those obtained only using sentiments, which had an accuracy of 725

2.54%. Therefore, understanding which topics the sentiments express is very useful in 726

forecasting the stock market. The topic-sentiment feature is better than the sentiment-only 727

feature. The accuracy obtained in this research was only 54.41%, which is unfavourable. 728

Furthermore, one of the weaknesses of their method is that the study only considered the 729

historical price and sentiments taken from social media. It was assumed that they kept 730

trying to find and integrate more factors that could affect stock prices to develop the 731

model with a higher accuracy. 732

A model using sentiment analysis to predict the stock market investment developed 733

by applying ANN for five companies—Apple, Google, Microsoft, Oracle and Facebook— 734

from 1\1\2015 to 22\2\2016 was proposed in [47]. Data were retrieved from Yahoo using 735

the Stock Twits website, which consists of five parameters: opening price, closing price, 736

high price, low price and volume. Validation graphs were used to present the errors using 737

MATLAB to compare the predicted price with the actual price. Four parameters—happy, 738

up, down and rejected—were used to determine if the data was positive, negative or neu- 739

tral. In this study, 75% of the data were used for training, 15% of the data were used for 740

testing and 10% of the data were used for validation. The LM algorithm and MSE for the 741

performance measure were utilised. It would be preferable to use more than one algo- 742

rithm to compare the results and select the best algorithm that outputs the least error. 743

744


15 | P a g e


After reading 45 research articles in the field of predicting stock market prices using 745

intelligent techniques, we found that many approaches can be employed to predict the 746

stock market price. Each of the reviewed studies proposed a method. Many algorithms can 747

be used for predictions; some studies even proposed a hybrid approach, combining more 748

than one algorithm in their model. While each algorithm differs, the following algorithms 749

are most frequently used to predict stock market prices: SVM, RNN, LSTM, ANN, KNN, 750

LR, and RF. 751

752

Each of the algorithms mentioned above has its advantages; while there is no “best” 753

or “worst” algorithm, there are factors that affect the performance of the algorithm. Thus, 754

before deciding on which algorithm to use, researchers should consider the factors that 755

may affect the outcome, such as: 756

• The size of the available dataset; 757

• Training time; 758

• Accuracy of the desired output; 759

• Number of features available. 760

761

Overall, many factors can affect the final prediction accuracy of each model. Those 762

factors include, but are not limited to, the ones mentioned above. The accuracy of each 763

model can also be improved by providing it with more training. Each of the studies pre- 764

sented in this paper can aid in deciding which algorithm is suitable for each situation. 765

766

3. Comparison and Analysis 767

In the previous section the reviewed articles were classified depending on the histor- 768

ical data, the sentiment data or both. However, some articles can be classified based on 769

several other criteria, including the algorithm used, whether the news was considered and 770

whether the context is local (in Saudi Arabia) or global. Table 1 summarises all previous 771

articles based on these criteria. 772

Table 1: Research comparison & analysis. 773

Authors Year Dataset Algorithm News Result Local/

Global

Selvamu-

thu et al. [3]

2019 Dataset:

1) Tick data of Reliance Pri-

vate Limited.

2) Data from Thomson Reu-

ter Eikon database.

No. Records: approximately

430,000 data points.

Section: Financial section.

Year: from 30 NOV 2017 to

11 JAN 2018.

ANN,

LM,Scaled

Conjugate

Gradient

and Bayes-

ian Regu-

larization

No The algorithms give a

precision of 99.9%.

Every one of LM, SCG,

and Bayesian Regulari-

zation the exactness

over the 15-min da-

taset drops to 96.2%,

97.0%, and 98.9% indi-

vidually.

Global


16 | P a g e


Zheng et al.

[4]

2017 Dataset: Alpha Vantage API

to access the time series data

of 82 randomly stocks

traded at New York Stock

Exchange.

No. Records: Not men-

tioned.

Section: Not mentioned.

Year: Not mentioned.

Logistic Re-

gression,

Bayesian

Network,

Simple

Neural

Network,

and SVM

No Accuracy rate 70% Global

Pathak et al.

[5]

2020 Dataset: Data of National

Stock Exchange of India

from Kaggle.


tioned.


Year: 2016-2017

RF, SVM,

KNN and

LR

No The best algorithm

was RF with an accu-

racy rate of 80.7%

Global

Alkhatib et

al.[6]

2013 Dataset: The stock infor-

mation of five randomly

choose companies recorded

on the Jordanian stock ex-

change.

No. Records: 1000 records.

Section: Educational, finan-

cial, electrical, investment

sections.

Year: from June 4, 2009, to

December 24, 2009.

KNN No The outcome presents

a decent indication

that the utilization of

data mining methods

could help decision-

makers at different

levels when using

kNN for data exami-

nation.

Global

Sen et al.

[7]

2020 Dataset: historical records of

NIFTY 50 indexes listed in

the National Stock Exchange

of India

No. Records: 50 indexes.


Year: 2008-2018.

CNN

No Mean 348.26 for one

week and mean 407.14

for two weeks.

Global

Cao et al.

[8]

2011 Dataset: Shanghai stock mar-

ket.

No. Records: not mentioned.


Year: 1999-2008.

ANN No The accuracy in 1999-

2002 with MAD

0.0107, MAPE 0.3125,

and MSE 0.2743.

Global

Liang et al.

[9]

2021 Dataset: Shanghai Stock Ex-

change composite index in

China.

No. Records: 1000 sample.


Year: not mentioned.

ANN (RBF

and MLFF)

No Performance of the

RBF is better than

MLFF because the er-

ror of RBF is much

smaller than the MLFF

Global


17 | P a g e


Abe et al.

[10]

2018 Dataset: MSCI Japan Index

components dataset.

No. Records: 319.


Year: from December 1990 to

November 2016.

NN, DNN,

SVR and

Random

Forest

No Deep neural networks

outperform shallow

neural networks in

general, and the top

networks also beat

typical machine learn-

ing models.

Global

Y.Lin et al.

[11]

2021 Dataset: Chinese stock mar-

ket.

No. Records: 65,000.


Year: 2000 - 2017

LR, SVM,

KNN, RF,

GBDT and

LSTM

No RF and GBDT have a

good predictive abil-

ity, LR prediction

needs to be improved,

KNN and SVM only fit

in some patterns,

LSTM model has more

unshown advantages.

Global

Ghania et

al. [12]

2019 Dataset: Yahoo Finance for

Amazon (AMZN) stock,

AAPL stock and GOOGLE

stock.

No. Records: January 2019 to

25 July 2019.

Section: Finance and tech-

nology sections.


LR, 3MMA,

and ES

No LR results were 24.31

3MMA results 21.08,

ES based on LR meth-

odology results were

16.62

Global

Kara et al.

[13]

2011 Dataset: ISE National Index's

daily closing price move-

ment.


tioned.


Year: From January 2, 1997,

through December 31, 2007

ANN and

SVM

No ANN model (75.74%)

SVM model (71.52%).

Global

Oliveira et

al. [14]

2013 Dataset: Economatica, Bra-

zil’s Central Bank,

BM&FBOVESPA and Thom-

son Reuters.

No. Records: 144 observa-

tions per month

Section: Banking Section.

Year: from January 2000 to

December 2011

ANN No Based on the set of er-

ror metrics with win-

dow size equal to 3, as

it presents a POCID

direction correct rate

of 93.62%, and MAPE

of 5.45%.

Global

Lozinskaia

et al. [15]

2017 Dataset: Russian MICEX

stock price index.


Section: financial.

Year: 2002–2016.

neural net-

work mod-

eling and

SVM.

No SVM algorithm pro-

duce better prediction

with

MSE=0.0009, RMSE=

0.0308, MAE= 0.0237,

MAPE= 0.0514,

R2=0.9728, and

cR2=0.5141.

Global


18 | P a g e


Kumar et

al.[16]

2011 Dataset: India TCS stock

market


Section: financial.


Genetic Al-

gorithm

(GA)

No Accuracy 99.87% Global

Soni et

al.[17]

2019 Dataset: the stock market

from the earlier year

No. of records: not men-

tioned

Section: not mentioned

Year: 2017

RF, SVM No the best algorithm for

predicting the market

cost of a stock is RF al-

gorithm

Global

Ebadati et

al.[18]

2018 Dataset: Apple, Pepsi, IBM,

McDonalds, and LG


tioned

Section: technology and food

Year: 3-Dec-2014 until

18-Sep-2015

GA with

ANN

No 99.42% SSE 88.75% re-

duction in time

Global

Khan at

al.[19]

2011 Dataset: ACI pharmaceutical

company

No of records: not men-

tioned

Section: pharmaceutical

Year: 31-08-2010 to 30-09-

2010

ANN No highest error rate was

3.28% and lowest was

0.12%

Global

Miah et

al.[20]

2015 Dataset: BEXIMCO Ltd


tioned

Section: conglomerate

Year: 13 days of January

2012

ANN with

fuzzy logic

No highest error rate was

4.8895% and lowest

was 0.3734%

Global

Asad et al.

[21]

2015 Dataset: Istanbul Stock ex-

change (ISE).

No. Records: 100 index.

Section: financial.


SVM.

No Accuracy is 70%. Global

Wang et al.

[22]

2021 Dataset: Alibaba, Pepsi- co,

VinGroup and Reliance.


Section: commercial.

Year:

Alibaba and Pepsi- co (2014-

2018)

VinGroup (2012-2015)

Reliance (2011-2019)

Linear re-

gression,

Polynomial

regression,

LSTM and

CONV1D

LSTM

No . CONV1D-LSTM al-

gorithm produce bet-

ter prediction with ac-

curcy of (54.17%)

based on Alibaba,

(51.56%) based on

Pepsi-co, (51.38%)

based on VinGroup

and (50.01%) based on

Reliance dataset.

Global

MOGHAR

et al.[23]

2020 Dataset:

New York Stock Exchange

NYSE (GOOGL and NKE).


Section: commercial and

trading

Year:

GOOGL (2004 -2019)

RNN No The result of the test

agrees that the model

can trace the evolution

of the rates of opening

prices for both assets.

Global


19 | P a g e


NKE (2010 -2019)

Lanbouri et

al. [24]

2020 Dataset: Kaggle

No. Records: 43148


Year: 2017 – 2018.

LSTM No Accuracy underneath

50%.

Global

Moghad-

dam et

al.[25]

2016 Dataset: NASDAQ stock

exchange


Section: industry

Year: 2015

ANN No The results show that

there is no difference

between the prediction

ability of the 4 and 8

prior working days as

input parameters.

Global

Oyewola et

al.[26]

2019

Nigerian Stock Exchange

(NSE) utilizing Guaranty

Trust Bank traversing

No. Records: not mention

Section: financial

Year: 2013-2018

LR, RF,

SVM, NN

No Top layer of RF has the

highest accuracy

MAE:0.4929

RMSE:0.6762

MSE:0.4573

MASE:0.2994

Global

Olatunji1 et

al. [27]

2013 Dataset: STC, SABIC and Al-

Rajhi bank

No. Records: 2130 from each

company.

Section: Energy, Telecom,

and banking sections.

Year:

STC: from 27th January 2003

to 22nd December 2010.

SABIC: from 6th January

1993 to 22nd December 2010.

ALRajhi Bank: from 9th Jan-

uary 1993 to 22nd December

2010.

ANN No Very low RMSE (Root

Mean Square Error)

down to 1.8174, very

low MAD (Mean Ab-

solute Deviation)

down to 18.2835, very

low MAPE (Mean Ab-

solute Percentage Er-

ror) down to 1.6476

and very high correla-

tion coefficient of up to

99.9%

Local

Alotaibi et

al. [28]

2018 Dataset: TADAWUL stock

market exchange and oil his-

torical prices


Section: industry section.

Year: 2003-2015.

ANN No Accuracy 75.7%. Local

Alturki et

al.[29]

2020 Dataset: Historical values of

SABIC, Alinma Bank, and

Alrajhi Bank.


Section: Banking section.

Year: from June 2018 to Au-

gust 2019

RNN No The model result was

satisfying contrasted

with obtained utilizing

the buy-and-hold trad-

ing method. So It is

good to think about

more factors to im-

prove the results.

Local

Alenezy et

al. [30]

2021 Dataset: Saudi Arabia stock

market (Tadawul), Saudi

Authority for Statistics, and

Saudi Central Bank

No. Records: 2026

Section: Banking section.

Combining

MODWT

functions

with the

ANFIS

model

No 3.3292731 ME Local


20 | P a g e


Year: 2011 – 2019

Jarrah et al.

[31]

2019 Dataset: Past data related to

stocks from the stock market

of Saudi (Tadawul)

No. Records: 130000


Year: 2011 – 2016

Combining

(DWT) and

(RNN)

No 0.15996 MAE - 0.03701

MSE - 0.19237 RMSE

Local

Alamro et

al. [32]

2019 Dataset: GDELT developed

by Google and TASI (Tad-

awul All Share Index)

No. Records: not mentioned


Year: not mentioned

LSTM No 0.59 MAE Local

Ghazanfar

et al.[33]

2017 Dataset: Karachi Stock Ex-

change (KSE) and Saudi

Stock Exchange (SSE).


Section: financial.

Year: Apr 2013 to Sep 2013.

SVM,

KNN, Ada-

boost, Na-

ïve Bayes,

Bayesian

Networks,

Multilayer

Perceptron

and RBF.

No Ada-boost, Multilayer

Perceptron and Bayes-

ian Network have

shown good results.

Both

Khan et al.

[34]

2020 Dataset: social media and

news (twitter).


Section: social media section.

Year: 2016-2018.

SVM Yes Accuracy 80.53%. Global

Al-Rubaiee

et al. [35]

2015 Dataset: all share sectors of

Saudi stock market that exist

in Tadawul website.

No. Records: 3335 records.

Section: social media section.

Year: from 17-3-2015 until

10-5-2015.

naive

bayes,

SVM, and

KNN

Yes The best results ob-

tained:

Recall for SVM: 95.71

Precision for KNN:

95.91.

Local

Moukalled

et al. [36] 2019 Dataset: AAPL, GOOGL,

AMZN and FB from Reuters

platform


Section: Trading.

Year: from January-01-2008

to December 31-2017

RNN,

DNN, SVM

and SVR

Yes Accuracy rate for SVM

APAPL: 82.91%

GOOGL:80.34%

AMZN:75.27%

FB:75%

Global

Patel et al.

[37]

2021 Dataset: India stock market

prices


Section: Financial

Year: 2020

ANN, Ran-

dom For-

est, KNN,

SVM, and

LSTM

Yes accuracy of KNN is

70%

accuracy of LSTM is

63%

accuracy of Random

Forest is 80%

Global


21 | P a g e


Giani et al .

[38]

2021 Dataset: stock price data was

taken from Yahoo Finance

and the Twitter data was ob-

tained from fol-

lowthehashtag


Section: Financial


NNs based

on LSTM

Yes accuracy of NNs is

76.14%.

Global

Kar-

lemstrand

et al. [39]

2021 Dataset: Historical stock val-

ues, technical indicators, and

Twitter attribute

No. Records: not mentioned

Section: social media, and

more

Year: historical stock price

data from Yahoo Finance for

up to 10 years

Neural net-

work with

LSTM

Yes 0.13 MSE Global

Pinheiro et

al. [40]

2017 Dataset:

1) Stock price data for all

S&P 500 from Thomson Reu-

ters Tick History

2) Financial news collected

from Reuters and Bloomberg


Section: Financial

Year: Financial news was

from October 2006 to No-

vember 2013

RNN and

NLP

Yes The outcomes suggest

that the utilization of

character-level embed-

dings is competitive

and promising with

more difficult models,

which use technical

pointers and occasion

extraction methods

with the news articles.

Global

Mehtab et

al. [41]

2019 Dataset: Daily historical data

of NIFTY 50 index during

January 2, 2015, till June 28,

2019, of India and twitter.


Section: not mentioned.

Year: 2015- 2019.

SVM, ANN

and LSTM

Yes Correlation 0.99,

MAPE 10.75, and

Matched Cases 80%.

Global

AR-

DIANTA et

al. [42]

2021 Dataset: Nine companies

from Yahoo Finance website:

Astra Agro Lestari (AALI,

Astra International (ASII),

Bank Central Asia (BBCA),

Merdeka Copper Gold

(MDKA), Pakuwon Jati

(PWON), Telekomunikasi

Indonesia (TLKM), Chandra

Asri Petrochemical (TPIA),

United Tractors (UNTR),

Unilever Indonesia (UNVR).

As well as Twitter accounts

like CNBC Indonesia.

No. Records: 124 transac-

tions per day.

Section:

Agriculture, Miscellaneous

Industry, Finance, Mining,

SVM and

NLP

Yes The average accuracy

rate was 65,33%.

Global


22 | P a g e


Property, Real Estate, and

Building Construction, Infra-

structure, Utilities, and

Transportation, Basic Indus-

try and Chemicals, Trade,

Service, and Investment, and

the Consumer Goods Indus-

try Sectors.

Year: from 6 July 2020 to 11

January 2021.

Kadiri et al.

[43]

2019 Dataset: BMCE BANK

No. Records: not men-

tioned.

Section: Financial

Year: 2016-2018

LSTM Yes RSM is equal to 2.04 Global

Nti et al.

[44]

2020 Dataset:

1) Authentic stock prices of

three companies (GCB,

MTNGH, and TOTAL)

2) Tweeter, web news

(myjoyonline.com, ghana-

web.com, and

graphic.com.gh), and post-

gathering.

3) Google trends

No. Records:

263 records from Google

trends

1581 post-gathering.

2184 tweets

Section: social media, Finan-

cial.

Year: from January 2010 to

September 2019.

ANN Yes accuracy of (49.4–52.95

%) based on Google

trends, (55.5–60.05 %)

based on Twitter,

(41.52–41.77 %) based

on a forum post,

(50.43–55.81 %) based

on web news, and

(70.66–77.12 %) based

on a combined dataset.

Global

Joshi et al.

[45]

2016 Dataset: Apple Inc. Com-

pany’s data.


Section: social media.

Year: from 1 Feb 2013 to 2

April 2016.

Sentiment

detection

algorithm

Yes Random Forest

worked admirably go-

ing from 88% to 92%

accuracy. The SVM ac-

curacy around 86%.

Naive Bayes algorithm

execution is around

83%.

Global

Nguyen et

al .[46]

2015 Dataset: Tow dataset, histor-

ical price from Yahoo Fi-

nance and mood infor-

mation from Twitter

No. Records: not men-

tioned.

Section: Financial.

Year: from July 23, 2012, to

July 19, 2013

SVM Yes accuracy is 54.41% Global


23 | P a g e


Khatri et al.

[47]

2016 Dataset: Apple, Google, Mi-

crosoft, Oracle, and Face-

book.


tioned.

Section: Technology

Year: from 1/1/2015 to

22/2/2016.

ANN Yes MSE for Apple: 0.14

MSE for Google: 0.27

MSE for Microsoft:

0.18

MSE for Oracle: 0.22

MSE for Facebook:

0.28

Global

774

Table 2 displays the most popular intelligent algorithms that were used in prior stud- 775

ies, as well as the outcomes of each algorithm. 776

Table 2: Common Intelligent Algorithms Applied. 777

Algorithm Papers that Applied it No. of Articles Result

KNN [5] [6] [11] [33] [35] [37] 6 Accuracy = 70%

Logistic Regression [4] [5] [11] [12] [26] 5 Accuracy = 78.6%

Simple Neural Net-

work

[4] 1 It is difficult to config-

ure and takes a while to

train

SVM [4] [5] [11] [13] [15] [17]

[21] [26] [33] [34] [35]

[36] [37] [41] [42]

15 Accuracy = 82.91%

Random Forest [5] [10] [11] [17] [26]

[37]

6 Accuracy = 80.7%

ANN [3] [8] [9] [13] [14] [19]

[20] [25] [27] [28] [37]

[41] [44] [47]

14 Accuracy = 77.12%

Bayesian Regulariza-

tion

[3] 1 Accuracy = 98.9%

CNN [7] 1 Mean for one week=

348.26

Mean for two weeks =

407.14

GBDT [11] 1 GBDT has a good pre-

dictive ability for short-

term prediction

RNN

Or

LSTM

[23] [29] [31] [36] [40]

[11] [22] [24] [32] [37]

[38] [39] [41] [43]

5

9

Accuracy = 81.3%

Combining MODWT

with ANFIS

[30] 1 3.3292731 ME

SVR [10] [36] [40] 3 Accuracy = 79.2%

NLP [40] [42] 2 Good performs and

simple in implementa-

tion

NN [10] [26] 2 Accuracy = 62.37%


24 | P a g e


DNN [10] [36] 2 Accuracy = 81.32%

3MMA [12] 1 21.08 average absolute

error

ES [12] 1 16 average absolute er-

ror

LM [3] 1 Accuracy = 96.2%

SCG [3] 1 Accuracy = 98.9%

Bayesian Network [4] [33] 2 It has shown a good re-

sult.

DWT [31] 1 Combining DWT and

RNN = 0.15996 MAE,

0.03701 MSE, 0.19237

RMSE

Sentiment Detection

Algorithm

[45] 1 Accuracy was going

from 88% to 92%

Naïve Bayes [33] [35] 2 Precision: 56.28

Recall: 73.59

Genetic Algorithm

(GA)

[16] 1 Accuracy 99.87%

Combining GA with

ANN [18] 1 99.42% SSE 88.75% reduc-

tion in time

Linear Regression [22] 1 Accuracy 55.97%

Polynomial Regres-

sion

[22] 1 Accuracy 52.81%

CONV1D LSTM [22] 1 Accuracy 54.17%

Ada-boost [33] 1 It has shown a good re-

sult.

Multilayer perceptron [33] 1 It has shown a good re-

sult.

RBF [33] 1 It has shown a good re-

sult.

778

Table 3 lists all the datasets that were used in previous studies, as well as the type of 779

dataset and the number of studies that have used it. 780

781

Table 3: Common Dataset Used. 782

Dataset/Company/Website No. of Research Historical Data / News

The Jordan steel company (JOST) [6] Historical Data

Irbid district electricity (IREL) [6] Historical Data

Arab international for education and investment

(AIEI)

[6] Historical Data

Arab financial investment (AFIN) [6] Historical Data

kaggork Stock Exchange [4] Historical Data

Data of National Stock Exchange of India from

Kaggle

[5] Historical Data

Tick data of Reliance Private Limited [3] Historical Data


25 | P a g e


Thomson Reuter Eikon database [3] Historical Data

Stock Exchange of India [7] [37] Historical Data

Shanghai stock market [8] Historical Data

Shanghai Stock Exchange composite index in

China

[9] Historical Data

MSCI Japan Index components dataset [10] Historical Data

Chinese stock market [11] Historical Data

Yahoo Finance for Amazon (AMZN) stock [12] [36] Historical Data

Apple stock [12] [36] [45] [47] Historical Data

GOOGLE stock [12] [23] [36] [47] Historical Data

ISE National Index's daily closing price move-

ment.

[13] Historical Data

Economatica, Brazil’s Central Bank,

BM&FBOVESPA and Thomson Reuters.


BEXIMCO Ltd [20] Historical Data

NKE [23] Historical Data

Kaggle [24] Historical Data

NASDAQ stock exchange [25] Historical Data

STC Dataset [27] Historical Data

SABIC Dataset [27] [29] Historical Data

AlRajhi Bank [27] [29] Historical Data

TADAWUL stock market exchange and oil histori-

cal prices


Alinma Bank [29] Historical Data

Saudi Central Bank [30] Historical Data

Saudi Authority for Statistics [30] Historical Data

Saudi Arabia stock market (Tadawul) [30] [31] Historical Data

GDELT developed by Google [32] Historical Data

TASI (Tadawul All Share Index) [32] Historical Data

Social media and news (twitter) [34] News

Twitter posts of all share sectors of Saudi stock

market that exist in Tadawul website.

[35] News

Facebook Shares [36] [47] News

stock price data was taken from Yahoo Finance

and the Twitter data was obtained from follow the

hashtag

[38] Historical Data and news

Historical stock values [39] Historical Data

Twitter attribute [39] News

Stock price data for all S&P 500 from Thomson

Reuters Tick History


Financial news collected from Reuters and Bloom-

berg

[40] News

Daily historical data of NIFTY 50 of India and

twitter.

[41] Historical Data and news


26 | P a g e


Astra Agro Lestari (AALI) from Yahoo finance

website


Astra International (ASII) from Yahoo finance

website


Bank Central Asia (BBCA) from Yahoo finance

website


Merdeka Cooprt Gold (MDKA) from Yahoo fi-

nance website


Pakuwon Jati(PWON) from Yahoo finance website [42] Historical Data

Telekomunikasi Indonesia (TLKM) from Yahoo fi-

nance website


Chandra Asri Petrochemical (TPIA) from Yahoo

finance website


United Tractors (UNTR) from Yahoo finance web-

site


Unilever Indonesia (UNVR) from Yahoo finance

website


Twitter account like CNBC Indonesia [42] News

BMCE BANK [43] Historical Data

Authentic stock prices of GCB [44] Historical Data

Authentic stock prices of MTNGH [44] Historical Data

Authentic stock prices of TOTAL [44] Historical Data

2184 tweets, web news (myjoyonline.com, ghana-

web.com, and graphic.com.gh), and 1581 post-

gathering.

[44] News

Google trends [44] Historical Data

Information from social networking sites [46] News

Oracle [47] Both

Microsoft [47] Both

Russian MICEX stock price index [15] Historical Data

Stock of TCS and Infosys [16] Historical Data

Apple [18] Historical Data

Pepsi [18] Historical Data

IBM [18] Historical Data

McDonalds [18] Historical Data

LG [18] Historical Data

ACI pharmaceutical company [19] Historical Data

Istanbul Stock exchange. [21] Historical Data

Alibaba [22] Historical Data

Pepsi- co [22] Historical Data

VinGroup and Reliance [22] Historical Data

Karachi Stock Exchange [33] Historical Data

Saudi Stock Exchange [33] Historical Data


27 | P a g e


Nigerian Stock Exchange (NSE) utilizing Guar-

anty Trust Bank traversing


783

4. Conclusions 784

People with different levels of knowledge and different backgrounds invest in the 785

stock market. In fact, the stock market offers great opportunities to make money. How- 786

ever, investing may involve great risks that might result in heavy losses. The availability 787

of obtaining accurate data in a timely manner could be the key factor in avoiding losses 788

and generating profits. Thanks to the 4th industrial revolution, several approaches, in- 789

cluding AI, ML and DM techniques, were developed for processing historical data in or- 790

der to generate formative knowledge that might help in decision making. These data pro- 791

cessing techniques (AI, DM and ML) have been proven to be effective in several domains, 792

and the stock market is no exception. 793

This article critically explored several intelligent techniques used to analyse historical 794

data and sentiment data that might affect stock market prices. Thus, it can serve as a start- 795

ing point for researchers and stock market experts who are interested in developing intel- 796

ligent techniques to predict stock prices. KNN, NN, SVM, Bayesian Network, DT and 797

other methods are among the different intelligent algorithms that have been used to create 798

intelligent stock market price prediction models. This article concluded that although 799

static intelligent algorithms have been shown to be effective, a robust model has not yet 800

been developed because these static approaches employ a static dataset and the produced 801

model cannot learn new knowledge after the training phase is completed. Therefore, be- 802

cause the stock market is a dynamic domain, not a static domain, stream DM can be as- 803

sumed to be a new possible direction for building prediction models that can learn new 804

knowledge as soon as it is generated. Furthermore, making use of fog and edge computing 805

will speed up the ability to collect, pre-process and analyse data in a timely manner. 806

Amongst other stream data mining algorithms, Hoeffeding Trees, Sliding Window and 807

Ensemble Learning can provide new research directions in order to build models that can 808

be described as strong, lifelong learning models. However, the availability of an accurate 809

training dataset could be one of the possible challenges. In the near future, it is important 810

to explore the capability of stream data mining techniques to develop intelligent stock 811

price prediction models. 812

References 813

[1] N. Klingler, “What Is an AI Model? Here’s what you need to know - viso.ai,” Viso.ai, Jul. 06, 2021. 814

https://viso.ai/deep-learning/ml-ai-models/ (accessed Nov. 19, 2021). 815

[2] S. Ray, “Commonly Used Machine Learning Algorithms | Data Science,” Analytics Vidhya, Sep. 09, 2017. 816

https://www.analyticsvidhya.com/blog/2017/09/common-machine-learning-algorithms/ (accessed Nov. 19, 817

2021). 818

[3] D. Selvamuthu, V. Kumar, and A. Mishra, “Indian stock market prediction using artificial neural networks on 819

tick data,” Financial Innovation, vol. 5, no. 1, p. 16, Dec. 2019, doi: 10.1186/s40854-019-0131-7. 820

[4] A. Zheng and J. Jin, “Using AI to Make Predictions on Stock Market,” Thesis, pp. 1–6, 2017, [Online]. Available: 821

http://cs229.stanford.edu/proj2017/final-reports/5212256.pdf. 822

[5] S. P. Ashwini Pathak, “Study of Machine learning Algorithms for Stock Market Prediction,” vol. 9, no. 06, 2020. 823

doi: 10.4018/978-1-7998-3645-2.ch007. 824

[6] K. Alkhatib, H. Najadat, I. Hmeidi, and M. K. A. Shatnawi, “Stock Price Prediction Using K-Nearest Neighbor 825

Algorithm,” International Journal of Business, Humanities and Technology, vol. 3, no. 3, pp. 32–44, 2013. 826

[7] J. Sen, S. Mehtab, and G. Nath, “Stock Price Prediction Using Deep Learning Models,” no. October, 2020. 827

http://cs229.stanford.edu/proj2017/final-reports/5212256.pdf


28 | P a g e


[8] Q. Cao, M. E. Parry, and K. B. Leggio, “The three-factor model and artificial neural networks: predicting stock 828

price movement in China,” Annals of Operations Research, vol. 185, no. 1, pp. 25–44, May 2011, doi: 10.1007/s10479- 829

009-0618-0. 830

[9] E. Theses and Y. Liang, “Stock Market Forecasting Based on Artificial Intelligence Technology,” 2021. 831

[10] M. Abe and H. Nakayama, “Deep Learning for Forecasting Stock Returns in the Cross-Section,” in Lecture Notes 832

in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 833

10937 LNAI, 2018, pp. 273–284. doi: 10.1007/978-3-319-93034-3_22. 834

[11] Y. Lin, S. Liu, H. Yang, and H. Wu, “Stock Trend Prediction Using Candlestick Charting and Ensemble Machine 835

Learning Techniques With a Novelty Feature Engineering Scheme,” IEEE Access, vol. 9, pp. 101433–101446, 2021, 836

doi: 10.1109/ACCESS.2021.3096825. 837

[12] M. A. and M. M. Mohammad Umer Ghania, “Stock Market Prediction Using Machine Learning Techniques,” in 838

Advances in Distributed Computing and Artificial Intelligence Journal, Sep. 2019, vol. 8, pp. 192–197. doi: 839

10.1109/ICAC49085.2019.9103381. 840

[13] Y. Kara, M. Acar Boyacioglu, and Ö. K. Baykan, “Predicting direction of stock price index movement using 841

artificial neural networks and support vector machines: The sample of the Istanbul Stock Exchange,” Expert 842

Systems with Applications, vol. 38, no. 5, pp. 5311–5319, May 2011, doi: 10.1016/j.eswa.2010.10.027. 843

[14] F. A. de Oliveira, C. N. Nobre, and L. E. Zárate, “Applying Artificial Neural Networks to prediction of stock 844

price and improvement of the directional prediction index – Case study of PETR4, Petrobras, Brazil,” Expert 845

Systems with Applications, vol. 40, no. 18, pp. 7596–7606, Dec. 2013, doi: 10.1016/j.eswa.2013.06.071. 846

[15] A. M. Lozinskaia and V. A. Zhemchuzhnikov, “MICEX INDEX FORECASTING: THE PREDICTIVE POWER OF 847

NEURAL NETWORK MODELING AND SUPPORT VECTOR MACHINE,” Вестник Пермского университета. 848

Серия «Экономика» = Perm University Herald. ECONOMY, vol. 12, no. 1, pp. 49–60, 2017, doi: 10.17072/1994- 849

9960-2017-1-49-60. 850

[16] R. Choudhry and K. Garg, “A Hybrid Machine Learning System for Stock Market Forecasting,” vol. 20, no. 1, 851

pp. 315–318, 2011. 852

[17] H. N. Shah, “Prediction of Stock Market Using Artificial Intelligence,” in 2019 IEEE 5th International Conference 853

for Convergence in Technology (I2CT), Mar. 2019, no. April, pp. 1–6. doi: 10.1109/I2CT45611.2019.9033776. 854

[18] O. M. Ebadati E. and M. Mortazavi T., “AN EFFICIENT HYBRID MACHINE LEARNING METHOD FOR TIME 855

SERIES STOCK MARKET FORECASTING,” Neural Network World, vol. 28, no. 1, pp. 41–55, 2018, doi: 856

10.14311/NNW.2018.28.003. 857

[19] Z. Haider Khan, T. Sharmin Alin, and A. Hussain, “Price Prediction of Share Market Using Artificial Neural 858

Network ‘ANN,’” International Journal of Computer Applications, vol. 22, no. 2, pp. 42–47, May 2011, doi: 859

10.5120/2552-3497. 860

[20] M. BadrulAlamMiah, “Price Prediction of Stock Market using Hybrid Model of Artificial Intelligence,” 861

International Journal of Computer Applications, vol. 114, no. 13, pp. 1–5, Mar. 2015, doi: 10.5120/20035-1134. 862

[21] M. Asad, “Optimized Stock market prediction using ensemble learning,” in 2015 9th International Conference on 863

Application of Information and Communication Technologies (AICT), Oct. 2015, no. October 2015, pp. 263–268. doi: 864

10.1109/ICAICT.2015.7338559. 865

[22] Q. Wang, K. Kang, Z. Zhang, and D. Cao, “Application of LSTM and CONV1D LSTM Network in Stock 866

Forecasting Model,” Artificial Intelligence Advances, vol. 3, no. 1, pp. 36–43, Apr. 2021, doi: 10.30564/aia.v3i1.2790. 867

[23] A. Moghar and M. Hamiche, “Stock Market Prediction Using LSTM Recurrent Neural Network,” Procedia 868

Computer Science, vol. 170, pp. 1168–1173, 2020, doi: 10.1016/j.procs.2020.03.049. 869

[24] Z. Lanbouri and S. Achchab, “Stock Market prediction on High frequency data using Long-Short Term Memory,” 870

Procedia Computer Science, vol. 175, no. 2019, pp. 603–608, 2020, doi: 10.1016/j.procs.2020.07.087. 871


29 | P a g e


[25] A. H. Moghaddam, M. H. Moghaddam, and M. Esfandyari, “Stock market index prediction using artificial 872

neural network,” Journal of Economics, Finance and Administrative Science, vol. 21, no. 41, pp. 89–93, Dec. 2016, doi: 873

10.1016/j.jefas.2016.07.002. 874

[26] D. O. Oyewola, E. G. Gbenga Dada, O. E. Olaoluwa, and K. A. Al-Mustapha, “Predicting Nigerian Stock Returns 875

using Technical Analysis and Machine Learning,” European Journal of Electrical Engineering and Computer Science, 876

vol. 3, no. 2, pp. 1–8, Mar. 2019, doi: 10.24018/ejece.2019.3.2.65. 877

[27] S. O. Olatunji, “Forecasting the Saudi Arabia Stock Prices Based on Artificial Neural Networks Model,” 878

International Journal of Intelligent Information Systems, vol. 2, no. 5, p. 77, 2013, doi: 10.11648/j.ijiis.20130205.12. 879

[28] T. Alotaibi et al., “Saudi Arabia Stock Market Prediction Using Neural Network,” International Journal on 880

Computer Science and Engineering, vol. 9, no. 2, pp. 62–70, Feb. 2018, doi: 10.21817/ijcse/2018/v10i2/181002024. 881

[29] F. A. Alturki and A. M. Aldughaiyem, “Trading Saudi Stock Market Shares using Multivariate Recurrent Neural 882

Network with a Long Short-term Memory Layer,” vol. 11, no. 9, pp. 522–528, 2020. 883

[30] A. H. Alenezy et al., “Forecasting Stock Market Volatility Using Hybrid of Adaptive Network of Fuzzy Inference 884

System and Wavelet Functions,” vol. 2021, 2021. 885

[31] M. Jarrah and N. Salim, “A Recurrent Neural Network and a Discrete Wavelet Transform to Predict the Saudi 886

Stock Price Trends,” International Journal of Advanced Computer Science and Applications, vol. 10, no. 4, 2019, doi: 887

10.14569/IJACSA.2019.0100418. 888

[32] R. Alamro, A. M. B, and A. A. B, Predicting Saudi Stock Market Index by Incorporating GDELT Using Multivariate 889

Time Series Modelling, vol. 1097. Cham: Springer International Publishing, 2019. doi: 10.1007/978-3-030-36365-9. 890

[33] M. A. Ghazanfar, S. A. Alahmari, asmeen F. Aldhafiri, A. Mustaqeem, M. Maqsood, and M. A. Azam, “Using 891

Machine Learning Classifiers to Predict Stock Exchange Index,” International Journal of Machine Learning and 892

Computing, vol. 7, no. 2, pp. 24–29, Apr. 2017, doi: 10.18178/ijmlc.2017.7.2.614. 893

[34] W. Khan, M. A. Ghazanfar, M. A. Azam, A. Karami, K. H. Alyoubi, and A. S. Alfakeeh, “Stock market prediction 894

using machine learning classifiers and social media, news,” Journal of Ambient Intelligence and Humanized 895

Computing, no. May, Mar. 2020, doi: 10.1007/s12652-020-01839-w. 896

[35] H. AL-Rubaiee, R. Qiu, and D. Li, “Analysis of the relationship between Saudi twitter posts and the Saudi stock 897

market,” in 2015 IEEE Seventh International Conference on Intelligent Computing and Information Systems (ICICIS), 898

Dec. 2015, pp. 660–665. doi: 10.1109/IntelCIS.2015.7397193. 899

[36] M. Moukalled and W. E. Mohamad, “Automated Stock Price Prediction Using Machine Learning,” 2019. 900

[37] A. Patel, D. Patel, I. Technology, I. Technology, S. Yadav, and I. Technology, “Prediction of stock market using 901

Artificial Intelligence,” 2021. 902

[38] K. Ranawat and S. Giani, “Artificial intelligence prediction of stock prices using social media,” pp. 1–13, Jan. 903

2021. 904

[39] R. Karlemstrand and E. Leckström, “Using Twitter Attribute Information to Predict Stock Prices,” May 2021, 905

[Online]. Available: http://arxiv.org/abs/2105.01402. 906

[40] S. Pinheiro and M. Dras, “Stock Market Prediction with Deep Learning : A Character-based Neural Language 907

Model for Event-based Trading,” 2017. 908

[41] S. Mehtab and J. Sen, “A Robust Predictive Model for Stock Price Prediction Using Deep Learning and Natural 909

Language Processing,” SSRN Electronic Journal, no. January, 2019, doi: 10.2139/ssrn.3502624. 910

[42] E. I. Ardyanta and H. Sari, “A Prediction of Stock Price Movements Using Support Vector Machines in Indonesia,” 911

vol. 8, no. 8, pp. 399–407, 2021, doi: 10.13106/jafeb.2021.vol8.no8.0399. 912

[43] A. K. Yamani, “SCHOOL OF SCIENCE AND ENGINEERING STOCK MARKET PREDICTIONS USING DEEP,” 913

no. April, 2019. 914


30 | P a g e


[44] I. K. Nti, A. F. Adekoya, and B. A. Weyori, “Predicting Stock Market Price Movement Using Sentiment Analysis : 915

Evidence From Ghana,” vol. 25, no. 1, pp. 33–42, 2020. 916

[45] K. Joshi, B. H. N, and J. Rao, “Stock Trend Prediction Using News Sentiment Analysis,” International Journal of 917

Computer Science and Information Technology, vol. 8, no. 3, pp. 67–76, Jun. 2016, doi: 10.5121/ijcsit.2016.8306. 918

[46] T. H. Nguyen, K. Shirai, and J. Velcin, “Sentiment analysis on social media for stock movement prediction,” 919

Expert Systems with Applications, vol. 42, no. 24, pp. 9603–9611, Dec. 2015, doi: 10.1016/j.eswa.2015.07.052. 920

[47] S. K. Khatri and A. Srivastava, “Using sentimental analysis in prediction of stock market investment,” in 2016 921

5th International Conference on Reliability, Infocom Technologies and Optimization (Trends and Future Directions) 922

(ICRITO), Sep. 2016, pp. 566–569. doi: 10.1109/ICRITO.2016.7785019. 923

924

925

CONFERENCES PAPERS


Sa’ah: Creative Eco-Friendly Mobile Application

That Encourages Living Sustainably.

Sara A. Alshaye, Aseel F. Almuhana, Shaima K. Alharbi, Jumanah M. Abudally, Nouf F. Alghamdi,

Dr.Gomathi Krishnasamy, Mona M. Altassan, Fatema S.Shaikh, Ghada M.Alrugaib

Department of Computer Information Systems

College of Computer Sciences & Information Technology

Imam Abdulrahman Bin Faisal University

Dammam, Saudi Arabia

Email {2180005522, 2180003558, 2180005994,

2180004556, 2180003582, gkrishna, maltassan,

gmalrugaib, fsshaikh}@iau.edu.sa

Abstract— In this time and age, with global warming and pollution rising by the minute, encouragement must be given to lessen their effect and protect our planet. Sa’ah app was developed with the hope of making the effort of reducing the waste that gets thrown out every day, easier. It targets three categories of waste reductions; reduce, reuse, and recycle. We reduce the waste by suggesting alternatives to reusing items and recycling them as well as the option of lending them out or simply donating them. Sa’ah app targets all the netizens by

providing a platform for easy access and real time communication between users for instructiveness and bigger support. It solves the problem of having to make more than one connection and more than one action to reach the desired result by giving access to all the solutions needed in just one platform.

Keywords: Mobile Application; System; Recycle; Donation; Borrow;

Image Processing, Eco-Friendly.

I. INTRODUCTION

Recently, this generation has been conscious and highly active in protecting the environment and are aware of the consequences of their waste. Out of the concept of the Arabic word “ سعه” we are aiming to make the effort - of reducing the waste that gets thrown out every day - easier with “Sa’ah” application. Sa’ah targets this generation's passion to live an eco-friendly life by providing three techniques to make use of the objects the user holds, a donation section, a recycling section, and a borrow section. The app aims to lessen the burden of the user when it comes to getting rid of unwanted objects by offering solutions according to the three categories the app holds. Promoting property value via recycling and reuse alternatives, as well as enabling innovative solutions and encouraging handcrafting users are redirected to a variety of charities and contribution sites.


II. LITERATURE REVIEW

This section involves several platforms supporting any of the three concepts that Sa’ah app includes, by briefly providing their services compared with Sa’ah Mobile application competitive features.

A. GoGreen

GoGreen [1] is a mobile application that helps the community and keeps the environment clean by providing scanning option that give the user brief information about the product material and will get information about where to throw the waste. as well as users will receive creative “do it yourself” ideas for the scanned product and eventually, the user will receive Eco-friendly suggested alternatives to the product to reduce causing harm to the environment.

The main drawback of GoGreen is that it doesn’t provide communication or a link between the users, as they cannot share their ideas.

B. OFree

OFree [2] is a mobile application developed to give, get or trade FREE stuff locally. It Saves the time by taking unwanted stuff like furniture, electronics, bikes, suitcases, clothing, but no food to a donation center or haggling over the price in an online selling forum. As well as users’ location will be detected to search for nearby traders that want the extra or unwanted stuff, users then can request an item to arrange for pick up. Users can communicate by private chat as they can make comments on the posts.

OFree is only focusing on the give and take concept which might not be a suitable option for all the users, as it doesn’t provide Recycling solutions.

C. Ea’arah Application

Ea’arah [3] is a free platform works in Saudi Arabia through which users can offer services, and items whether for free or not to other people. The user must specify the living area to view the products available in his area. Users can also use the filter to directly identify the product or to search for customized things, determine their need, and contact the seller or product owner to get it through a private chat as they can make comments in the displayed advertisement.

The app’s location span is limited only to Eastern province which could be a drawback for the app, as it’s only supporting the concept of Borrowing.

D. RecycleCoach

RecycleCoach [4] is a mobile application that teaches the users how to recycle in a correct way and how to sort recyclable items. The app begins by conducting a survey to test and correct the knowledge of the user regarding recycling. It offers a search tool for users to search if an item can be recycled or not, and if yes, how to recycle it. The app also contains a calendar for scheduling recycling days and setting reminders. As it contains a discovery zone with a blog that contains information.

The app is not providing creative recycling solutions and the products cannot be scanned, which is considered a main drawback for the app.

E. FreeSpot

FreeSpot [5] is a mobile application that provides a crowd-sharing platform to connect people to free food, clothing, furniture, events and more that would otherwise go

to waste. This effectively reduces both the amount of food and toxic waste while supplementing meals and material necessities for communities in need. Users can make claims on food, events, furniture and more by opening a private chat with the item's poster as they can Join the campaigns created by nonprofits in the area seeking free items to be donated from the communities they support.

What lacks FreeSpot is that users must give and take hand by hand rather than directing the user to a charity.

This paper proposes an application system Combines three Eco-friendly options (Recycling, Donations and Borrowing). Table 1 as shown below summarizes the main similarities and dissimilarities between the proposed system and discussed systems including users of the system.

Table 1: characteristics of common applications compared to Sa’ah.

III. MOTIVATION

Sa’ah Mobile Application solves many environmental issues and provides solutions that could protect our environment. The main objective is to provide easy and creative ways of recycling and reusing, as well as facilitate the donation and borrowing process by easy directing or providing communication between users who have the same interest. We conducted a survey to know how people are conscious and caring about Sa’ah app concept, the survey filled by 245 users, and we received positive feedback.

The role of information systems is vital in this project as it supports decision making and communication coordination and control ad analysis, so it is applied in every stage of the project.

Application aims and objectives are:

1. Raising environmental responsibility and awareness toward excessive waste by providing 3 ways of item utilization: donation, borrowing, and recycling.

2. Reinforcing the concept of property valuation by providing recycling and reusing solutions.

3. Helping lenders or donators to have more space and get

Characteristics Ofree FreeSpot GoGreen RecycleCoach Ea’arah Sa’ah

Private Chat ✓ ✓

✓

Support

Payment

✓

✓

Shared

Dashboard ✓ ✓

✓ ✓ ✓

Free Trading ✓ ✓

✓ ✓

Products

Scanning

Option

✓

✓

Creative

Recycling

Solutions

✓

supports

Donating

✓

✓

Combines

(Recycling,

Donations and

Borrowing)

✓


rid of overstocked items, in return saving borrowers’ money by borrowing and donating solutions.

4. Providing creative solutions and reinforcing handcrafting by suggesting different usages for the scanned item using image processing where the image is the input, and the alternative recycling solutions are the output.

5. Allowing users to access the borrow and recycle stream feed with items’

specifications and return policy.

6. Allowing users to comment and share ideas of any feed in the stream.

7. Directing users to several charities and donation platforms.

8. Providing time and date schedule and borrowing period for the borrowed items

IV. METHODOLOGY

The choice of tools used to implement Sa’ah application has been selected according to what will give the user the best experience. Here are the detailed factors that lead to the choice of programming language and database:

A. flutter Programming language

Flutter program allows to produce the application in two operating systems (Android and IOS) using only one programming language and one codebase and provide the system to a wider audience.

B. Firebase Database

The decisions to use a Firebase real-time database were made according to the demand of real-time and rapid response for the user to have the best experience as well as for the purpose of storing and retrieving users’ information to ensure authentication and other functionality securely.

V. IMPLEMENTATION

Sa’ah is a mobile application that runs on IOS and Android operating systems. The system implemented to serve society and the environment by providing alternative methods to make use of any object the user has, a donation section, a recycling section, and a borrow section.

The main users of the application are admin and user. Each of the main users has different functionalities in respect of their need and privilege. A set of user interfaces have been developed allowing the admin to add, delete, update accounts of the users, as well as block keywords that must not be used within the users for ethical reasons and for better service quality. In addition, the admin can add, delete, deny, and approve the posts with the help of the system moderator.

On the other hand, users of Sa’ah application can interact and benefit from services of the application. The set of user interfaces are described below.

• Dashboard interface

The community communicate through a shared dashboard that Contains post and articles and user can navigate to the multiple pages in the app. • Scan interface

Users can use the camera to take a photo of an item or to

choose a photo from the photo library to be scanned. After scanning two options are available, whether to show relative recycling ideas or to show a suitable donation center that accept the scanned item.

• Add New post Interfaces

Users can add a new post and can turn on the camera or choose a photo from the library to upload object’s picture. The user will be provided to choose from three options, ‘Recycling’ option where the post will be posted on ‘Recycle’ dashboard, ‘Donating’ option where the user will be directed to ‘Donate’ interface that includes a set of charities. And if the user chooses the ‘Borrow’ option, then the post will be posted on ‘Borrow’ dashboard.

• Recycle Interfaces

Recycle interface includes the posts and articles that are only related to Recycling. Users can interact with others as they can share posts, save posts, and make comments.

• Donate Interface

Donate interface includes the charities that can accept the items chosen. Users can read a brief about the charity and will be able to contact the charities list shown easily.

• Borrow Interfaces

Borrow interface includes the posts that are only related to Borrowing. Users can share posts, make comments, save posts, mark the post status, and contact traders to loan the items. As well as ‘Borrow’ dashboard can be filtered and sorted according to the user’s preferences.

The figure below shows the System flow diagram of Sa’ah mobile application. See figure1.

Figure 1: System Diagram of Sa’ah application.


VI. CONCLUSION

This paper represents Sa’ah app system that provides creative eco-friendly solutions and encourages living a sustainable life. The goal of Sa’ah app is to protect the environment and to help the people live a minimalistic lifestyle that's suitable for the environment. with the harness of technology Sa’ah app aims to serve society and Promote property value via recycling/reuse, borrowing/lending and donating.

VII. Appendix

This survey was conducted on 245 surveyors. it shows the need

for Sa’ah app and shows how acknowledged people are

regarding sustainability and minimalism concepts. The results

of the survey are shown below as Pie chart.

Question 1 (Have you ever tried living an eco-friendly, sustainable life?)

The answers were mostly "Kind of " followed by "Yes " meaning the majority try to be environmentally responsible. See Fig2.

Figure 2 Question 1

Question 2 (Do you suffer from overstocked belongings?)

The answers were mostly "Yes", followed by "No " and "Kind of " having similar percentages. See Fig3.

Figure 3 Question 2

Question 3 (Are you familiar with the concept of “minimalism”?)

Figure 4 Question 3

Question 4 (Have you ever considered donating or borrowing?)

The answers were mostly "Yes" indicating most people are open to the concept of donation ad borrowing. See Fig5.

Figure 5 Question 4

Question 5 (On a scale from 1 to 5 how much do you think alternative ideas of recycling objects will help in protecting the environment?)

The answers were mostly "high effect" meaning that majority agree that Recycling can and will benefit the environment. See Fig6.

Figure 6 Question 5

Question 6 (Have you ever thrown an item away because you didn't know how to reuse or recycle it?)

The answers were mostly "Yes", this gives more reason for why the application should exist and raise awareness on recycling and its alternatives. See Fig7.

The answers were mostly "No" followed by "Yes" meaning that most people don’t know about this concept and the app should be a way to provide such knowledge. See Fig4


Figure 7 Question 6

Question 7 (Do you think people have a background on how to recycle?)

The answers were mostly "Kind of" meaning that most people don’t have enough knowledge on how to live an eco-friendly lifestyle. See Fig8.

Figure 8 Question 7

Question 8 (Do you think borrowing items over the internet in Saudi Arabia can be possible?)

The answers were mostly "Yes" indicating that people are open to the idea and are not totally opposed to it as a very little percentage was "No". See Fig 9.

Figure 9 Question 8

Question 9 (Do you prefer to donate through applications or to go to charities themselves?)

The answers were mostly "Through Application" meaning that most people prefer to use apps to donate which correlates with our applications purpose. See Fig 10.

Figure 10 Question 9

Question 10 (Which of these three appeals to you the most?)

The answer mostly "Donating" meaning the majority tend to donate, followed by recycling, and last was borrowing, meaning people have still yet to practice the idea of lending and borrowing. See Fig 11.

Figure 11 Question10

Figure 12 Question 11

Question 11 (On a scale from 1 to 5 what’s the possibility of downloading an application that can help on the three parts above?)

The answers mostly "High possibility" according to the graph, most people find a high possibility of downloading an application that combines all the options. See Fig 12.


6 | P a g e

Published In: 2022 8th International Conference on Advanced Computing and Communication Systems (ICACCS) eCF Paper Id: 374899

REFERENCES

[1] GoGreen is an IOS application that monitors your carbon

footprint. download Today to discover your impact.

GoGreen. (n.d.). Retrieved February 28, 2022, from

http://www.gogreenapp.org/

[2] Gfreeapp is a technology related blog that covers the latest

news. Ofreeapp.com. (2021, December 8). Retrieved

February 28, 2022, from https://ofreeapp.com/

[3] Abdulgader, R. (2018, September 27). إعارة. App Store.

Retrieved February 28, 2022, from

https://apps.apple.com/sa/app/%D8%A5%D8%B9%D8%A7%D

8%B1%D8%A9/id1434242614

[4] Save your municipality's recycling program. Recycle Coach.

(n.d.). Retrieved February 28, 2022, from

https://recyclecoach.com/

[5] Co., F. S. (2019, September 1). FreeSpot Co.. App Store.

Retrieved February 28, 2022, from

https://apps.apple.com/sa/app/freespot-co/id1477121194

[6] Resources, Conservation and Recycling | Vol 136, Pages 1-488

(September 2018) | ScienceDirect.com by Elsevier. (2022).

Retrieved 14 February 2022, from

https://www.sciencedirect.com/journal/resources-conservation-

and-recycling/vol/136/suppl/C

[7] Hornik, J., Cherian, J., Madansky, M., & Narayana, C. (1995).

Determinants of recycling behavior: A synthesis of research results.

The Journal Of Socio-Economics, 24(1), 105-127. doi:

10.1016/1053-5357(95)90032-2

ACKNOWLEDGMENT

We would like to start with expressing all our special thanks and gratitude to Allah for his facilitate and patience he gave us during this project and to enlighten our minds with such things that could help a whole community.

The team members would thank the project supervisor Dr. Gomathi Krishna for guidance and continuous supervision and the Computer Information Systems department and the College of Computer Sciences & Information Technology for providing the necessary support and resource.

As a cooperative team member, we would like to express our gratitude to each other for the cooperative and the hard work that been done, and for all the time that have been spent to prepare this project.

[8] Jenkins, R., Molesworth, M., & Scullion, R. (2014). The

messy social lives of objects: Inter-personal borrowing and

the ambiguity of possession and ownership. Journal Of

Consumer Behaviour, 13(2), 131-139. doi: 10.1002/cb.1469

[9] Hopewell, J., Dvorak, R., & Kosior, E. (2009).

Plastics recycling: Challenges and opportunities.

Philosophical Transactions of the Royal Society B: Biological

Sciences, 364(1526), 2115–2126.

https://doi.org/10.1098/rstb.2008.0311

[10] Cohen, S. (2017). Understanding the Sustainable

Lifestyle, 1–4. Retrieved from

https://www.researchgate.net/profile/Steven-Cohen-

4/publication/325273718_Understanding_the_Sustainable_

Lifestyle/links/5b032498aca2720ba098fef6/Understanding-

the-Sustainable-Lifestyle.pdf?origin=publication_detail.

[11] A. Martin-Woodhead, “Limited, considered and

sustainable consumption: The (non)consumption practices

of UK minimalists,” Journal of Consumer Culture, p.

146954052110396, 2021.

[12] A. Bartl, “Moving from recycling to waste prevention:

A review of barriers and enables,” Waste Management

& Research: The Journal for a Sustainable Circular

Economy, vol. 32, no. 9_suppl, pp. 3–18, 2014

https://www.sciencedirect.com/journal/resources-conservation-and-recycling/vol/136/suppl/C

https://www.sciencedirect.com/journal/resources-conservation-and-recycling/vol/136/suppl/C

https://doi.org/10.1098/rstb.2008.0311


1 | P a g e

Published In: 5th International Conference on Multi-Disciplinary Research Studies and Education (ICMDRSE-2022)

Aknaf Website: Interactive Website to Automate the

Institution’s Work

1st Bedour Aljindan

Department of Computer Information

System,

College of Computer Science and

Information Technology,

Imam Abdulrahman Bin Faisal

University,

P.O Box 1982

Dammam 31441, Saudi Arabia

[email protected]

4th Renad Alrabeea


System,




University,

P.O Box 1982


[email protected]

2nd Leena Kanadiley


System,




University,

P.O Box 1982


[email protected]

5th Wafa Alsharekh


System,




University,

P.O Box 1982


[email protected]

3rd Renad Alghamdi


System,




University,

P.O Box 1982


[email protected]

Abstract— Technology is occupying a big part of our lives,

which could become an essential part. All institutions these days

need the use of technology to reach their peak and succeed. yet

unfortunately, there is a large percentage of educational centers

and consulting institutions that do not have a digital strategy,

despite the advantages they could gain from technology. A

meeting was held with the administrators of the Aknaf

institution to discuss the problems they are facing. This project

aims to develop a website for Aknaf institution, which is

interesting because Aknaf still performing all their task

manually without any help of technology. Current commercially

available websites do not cater for all requirements. This paper

will help Aknaf institution to make things easier and to achieve

its desired goal to reach the largest segment of beneficiaries and

save operation costs for the institution. However, such a website

can be extended to work in other consultation institutions.

Keywords—Website, Aknaf, QR code, Technology

I. INTRODUCTION

Technology is becoming an essential factor to

organizations; we are living in the digital age that no sector

can deny the fact it is crucial for their organization success.

Digital technologies affect how a company interacts with

consumers and partners, transforming internal processes and

creating opportunities for discovering and implementing new

techniques for company growth [1].

The goal of Aknaf website is to automate the work for the

institution to reach the largest segment of beneficiaries and

save operation costs. The system has three different levels of

users: the admin, the clerks, and the clients. The admin on the

website is the institution manager. The admin can manage

users, programs, digital library, and consultation

appointment. Also, he/she can view and generate reports and

approve payment receipts. The clerk can manage programs,

consultation appointments and reply to the clients. Also,

he/she can generate QR code for the program’s attendance.

The clients can view programs and book an appointment for

a consultation. Only registered clients can view E-books in

the digital library and make the payment. All users can

customize the settings of the website such as choosing the

website color and modifying profile.

The website focusses on the problem associated with fulfilling institution work and reaching clients. It considers in detail the issues that arise due to the lack of a website that brings together all the operations. Thus, it causes the loss of time and effort of the institution’s administrators and the difficulty of communicating with clients.

II. PROBLEM STATEMENT

The aim is to point out the problem to function better.

Aknaf has very primitive ways regarding fulfilling their work

and reaching clients. They still facing difficulty with

documentation, registration, appointment scheduling since

regular activities still manually accomplished. However,

delivering their voices to the public is one of the main issues

that they are facing, they spend many hours recording the

client’s number and sending a message individually which is

a very untechnical way. Moreover, Aknaf needs to analyze


2 | P a g e


their client’s attitude and preferences to increase their profit

by gaining more audiences, which they need reports to help

them make decisions.

Lack of a database prevents Aknaf from reaching their goals and objectives. Creating a database that records the client’s information will make the work easier and more efficient. Moreover, having an official website for the institution will make operations much easier, faster, and attract more clients. the goal is to solve the institution’s problems and make their operations go faster and smoother.

III. LITERATURE REVIEW

This section defines the similar websites that are used in

charities, training centers, and training courses. It discusses

the similarities and differences between some of these

systems, and the proposed system is described.

A. Saudi Cancer Society Website

Saudi cancer society encourages scientific research to stop

and identify the causes of cancer in the Kingdom. The website

provides many features that benefit visitors such as activities

and events, reports, a questionnaire, an introductory film about

the association, premium partners, about the association,

assembly Services, detection centers, translated books, Media

Center, and a staff portal. The website provides appointment

scheduling for earlier cancer detection. First, it requires

answering a few questions such as name, national ID. after

that the user can schedule an appointment [2].

B. Unit Success Skills Training Center Website

Success Skills Training Center is specialized in the field of human development, training, building, and development of human cadres, with its capabilities, equipment, capabilities, expertise, and trainers capable of providing the highest levels of training, is considered one of the most important houses of expertise in the Arab world. Success Skills Training Center provides many courses to teach people how to be successful, they have a big website that includes all the advanced sections and functions such as, who we are, news, articles, library, training bag, our clients, success partner, courses schedule, reports, and calls us. The user is able to view all the available courses and purchase what he needs which they call training bags. Moreover, the Success Skills Training Center display continuous and brief update about their report which include the numbers of the reports they have, the numbers of images, the number of views, and the number of prints, the user can search for a certain report by filling in the required information [3].

C. Droob Course website

Droob is a national platform that is an infinitive of the Saudi Human Resources Development Fund. Droob website allows users to see what courses are available and register at any one of them. No payment is required since all courses are offered for free. The website has an Ad bar and a FAQ page. Registration on the website is done using the national Saudi ID. The website offers courses in many departments such as IT, languages, finance, health, and others. Contacting Droop is done using either filling out a form or contacting them during working hours on their official number [4].

D. Mnar Website

Mnar is a unified electronic platform that allows the

trainee to review training courses in the Kingdom of Saudi

Arabia under one roof, as well as enabling him to view all data

related to training institutes and trainers [5]. The platform

displays the course and important information such as city,

time, price, language, and target group. One of the features is

that customers' comments on the courses can be viewed and

interacted with. It also allows customers to share courses via

social networking sites. There is also a help box on the site

where customers can leave a message for the organization,

whether for inquiries or technical support. When registering

with the platform, customers will receive the latest training

courses via their e-mail.

E. Udemy Website

Udemy is a platform that offers both paid and free courses,

as well as the ability for instructors to create online courses

on their preferred topics. Furthermore, it assists organizations

of all sizes and types in preparing for the path ahead,

wherever it may lead. Courses are available in a wide range

of categories, including business and entrepreneurship,

academics, arts, health, and fitness, language, IT & Software,

and each course has a rating and reviews. When a user

registers on Udemy, a dashboard appears that contains all

training courses for which the user has registered. The Udemy

website saves all purchases made by users and provides a

receipt for each purchase [6].

This paper proposes a website to act as a middle link

between clients and admins. Table 1 as shown below

summarizes the main similarities and dissimilarities between

the proposed system and discussed systems including users

of the system.


3 | P a g e


Table 1: Summary of similarities and dissimilarities

IV. MOTIVATTION

There are several motivations that have encouraged the team members to develop Aknaf Website:

• The website will be a community service since the team members will develop and deliver the website for free.

• The system will be used in real-time under the name of the team.

• It aims to solve a real-life problem and will help Aknaf enhance its performance because it solves a problem that exists in real life.

V. OBJECTIVES

The objective of the project is to create an official

platform for the institution to use. Having a website is

essential for Aknaf to reach its goal and a wider range of

users. By creating a database and a website hoping to help

Aknaf institution work easier, faster, and more efficiently.

• Help Aknaf institution to reach a wide range of

beneficiaries.

• Ease the process of registration.

• Ease in report generation.

• Ease in programs and consultations scheduling.

• Ease in taking attendance of programs through QR code.

• Ease of displaying and selling the institution's

publications, including books and articles.

VI. METHODOLOGY

The tools used to implement Aknaf Website were chosen based on their compatibility with system requirements. Here are the specific considerations that influence programming language, server, and database selection:

A. PHP Programming Language:

PHP is considered the first server-side language that could be embedded into HTML. It allows website pages to load faster. Also, it may be used on any primary operating system.

B. XAMPP Server:

XAMPP enables a local host or server to test its website by computers or laptops before launching it on the main server. It is a simple, flexible, and lightweight tool that can facilitate website testing and development process [7].

C. MySQL Database:

MySQL controls how quickly things load on a website

and how quickly that stored data can be accessed. It has a

direct effect on website performance, making it an essential

part of web design [8].

VII. IMPLEMNTATION

Aknaf is a website for training and consultation that runs

on the iOS and Windows operating systems. This website will

help the Aknaf institution to automate manual activities, and

it will also help to reach a wide range of beneficiaries.

Moreover, it helps the admin to manage the process of

booking consultation appointments for the clients, and the

clerk to manage the attendance of programs through

generating QR codes, the admin can make a better decision

for programs, and consultation appointments by generating

reports. it helps both admin and clerk in scheduling the

program.

Different users of the system have different levels of

authority. Fig.1 shows the main three users of the system and

their functionality. The following section will specify some

features and functions for admin, clerk, and client.

Sys

tem

System users Similarity Dissimilarity

Saud

i C

ance

r

Soci

ety

Patients, Staff

Both systems

provide an appointment

scheduling

- There is no

dashboard for the

users - There is no QR

scanning for

attendance

Succ

ess

skil

ls

Trainers, and clients

Both systems generate reports.

- There is no QR

scanning for

attendance

Dro

ob

Cou

rse

clients

Both systems allow users to

view what

programs are available and

register at any one

of them.

- There is no QR

scanning for attendance

- The system does

not allow viewing customer

comments

Mna

r W

ebsi

te

Clients

Both view clients'

comments.

- There is no

digital library

- There is no QR scanning for

attendance

Ud

emy

Instructors, and

learners

Both offer paid

and free programs.

- There is no QR

scanning for attendance


4 | P a g e


Fig. 1: Aknaf Use Case Diagram

Fig.2: Sequence diagram (generate QR code)

Fig.3: Activity diagram (rating/add comments)

• Schedule program

In the admin interface, the admin will be able to add

a program and schedule it by sign-in into the dashboard

and then clicking on add program, after that she/he will

be able to fill in all the required information about this

program.

• Generate QR code

QR Codes will be generated for clients to check-in

for the offline and online attending programs. The user

who is responsible is the clerk will sign into the account

and choose a program then click on generate QR it will

be generated by the system. Fig. 2 shows demonstrate the

interaction and relationship between the components.

• Digital library

The clients can view the digital library that contains

articles and books and can choose either to view books

or articles. Books will be available only when a client is

registered on the website, they can also purchase a book.

• Rating/add comments

The website will provide feature to rate e-books/articles

and comments, customer can provide feedback to Aknaf team

by leaving a comment or rate e-books/articles. Fig. 3 shows

the workflow.

• Generate report

Generate report interface is for the admin. The system will generate report with information about the programs, consultation appointment, block users. Fig. 4 shows demonstrate the interaction and relationship between the components.


5 | P a g e


Fig.4: Sequence diagram (generate report)

Fig.6: Activity diagram (color change)

• Managing consultation appointment

In the admin interface, the admin will be able to

manage the appointment times and reasons. The admin will receive the suitable time slots for the consultant then the admin will enter the time slots in the appointment sections.

• Payment management system

The clients must register to the website to be able to enroll in Programs or book consolation appointments. Moreover, they can choose a suitable payment method and then fill in the payment information. Finally, pressing enter to validate the credit card if the card is Rejected the website will display a rejected message to the client if approved the website will display a confirmation message, and the client Program will be automatically added to the client Program list. Fig.5 shows demonstrate the interaction and relationship between the components when the client wants to book an appointment.

Fig.5: Sequence diagram (payment management system)

• Color change

Now a days there is awareness of the color-blind issue, so the website will provide the user the ability to change the theme color of the website. The user will click on the settings, then they will have a color picker option. The users can choose a color that is suitable for them. Fig. 6 shows the workflow.


6 | P a g e


VIII. RESULTS

The project began with meeting with the administrators

that need an online system to solve problems and reach out to

a bigger audience. The aim of the project was to develop a

website based on Aknaf institution requests that will allow its

clients to schedule appointments and programs online rather

than the old manual way and to reach more clients than they

currently have. The admin has important functionalities like

generating reports and QR codes that will help the company

achieve its desired goals faster. The team members

successfully planned and managed the project which led to a

successful implementation of the system proposal. The team

has successfully analyzed all requirements and functionalities

of the system and prepared everything needed for the project

implementation.

IX. DISCUSSIONS

In the line with previous studies, after analyzing all

websites we had noticed there isn’t a website that has the QR

scanning functionality which is the main function of Aknaf

that enables the client to scan the QR code by their phones

and confirms their attendance to the course. All websites

found are very well done and easy to use. Some websites have

more functionality than others, while some lack many key

features such as appointment reminders, report generating,

ratings and reviews.

X. CONCLUSION

This paper represents the work required to construct the

Aknaf Website that facilitates communication for the Aknaf

institution administrators and their clients. However, it can

certainly be applied to other institutions. Therefore, the goal

was to employ the available technology to develop websites

that serve institutions for their growth and attract more

audiences. Clients can use the system to view and book the

institution's multiple programs and can also browse the

institution's digital library. The system also facilitated the

process of booking consultation appointments. On the other

hand, this system gives admins to keep track of processes

such as payment approval. And the capability to manage

users, programs, consultation appointments, and digital

library.

The recommended future work is to develop a mobile

application to reach out to a wider range of clients. Also,

adding features like live chat with the clients to make

communication easier.

ACKNOWLEDGMENT

Firstly, we are thankful to Almighty Allah for his grace, his continued bounty, and for getting us accepted into this scientific conference.

It is the team members' pleasure to acknowledge the support and assistance received from the Computer Information Systems department and the College of Computer Sciences & Information Technology. Also, we would like to acknowledge the help of the Aknaf institutions in providing the team with the necessary information.

Additionally, the members of the team would like to express their gratitude, appreciation, and respect for the hard

work, effort, time, and sincerity they have displayed in this project.

REFERENCES

[1] T. Averina et al, "Impact of digital technologies on the company’s business model," E3S Web of Conferences, vol. 244, 2021. Available: https://www.e3s-conferences.org/articles/e3sconf/pdf/2021/20/e3sconf_emmft2020_10002.pdf [Accessed 05 October 2021].

[2] Saudicancer.org. 2021. الجمعية السعودية لمكافحة السرطان. [online] Available at: <https://www.saudicancer.org/> [Accessed 1 October 2021].

[3] Skills, s., 2021. details. [online] Sst5.com. Available at:

<https://sst5.com/> [Accessed 6 October 2021].

دروب" [4] :Doroob.sa, 2021. [Online]. Available ,"برنامج https://doroob.sa/ar/. [Accessed: 06- Oct- 2021].

[5] Mnar.sa. 2021. منار. [online] Available at:

<https://mnar.sa/> [Accessed 6 October 2021].

[6] "Online Courses - Learn Anything, On Your Schedule | Udemy", Udemy, 2021. [Online]. Available: https://www.udemy.com/. [Accessed: 06- Oct- 2021].

[7] P. Kumari and R. Nandal, "A Research Paper OnWebsite Development OptimizationUsing Xampp/PHP," International Journal of Advanced Research in Computer Science, vol. 8, (5), pp. 1231-1235, 2017. Available: https://www.proquest.com/docview/1912631847/9E98A185D4C34882PQ/1?accountid=136546 [Accessed: 06-Nov-2021]

[8] Mysql.com. n.d. Why MySQL?. [online] Available at: <https://www.mysql.com/why-mysql/> [Accessed 8 November 2021].

https://www.e3s-conferences.org/articles/e3sconf/pdf/2021/20/e3sconf_emmft2020_10002.pdf



https://www.saudicancer.org/

https://sst5.com/

https://mnar.sa/

https://www.proquest.com/docview/1912631847/9E98A185D4C34882PQ/1?accountid=136546

https://www.proquest.com/docview/1912631847/9E98A185D4C34882PQ/1?accountid=136546


1 | P a g e

Published In: Conference 2d ASIANCON 2022 Pune, India

Flourish: Requirements and Design of an Android

Application Prototype for Various Symptoms

Management in ADHD Patients

1st Sardar Zafar Iqbal


Systems, College of Computer Science

and Information Technology, Imam

Abdulrahman Bin Faisal University,

P.O Box 1982


[email protected]

4th Lubna Alghamdi





P.O Box 1982


[email protected]

7th Hina Gull





P.O Box 1982


[email protected]

2nd Aroob Alkarni





P.O Box 1982


[email protected]

5th Monera Almokainzi





P.O Box 1982


[email protected]

8th Ruba Alsalah





P.O Box 1982


[email protected]

3rd Jumana Aleleyo





P.O Box 1982


[email protected]

6th Rayanh Alyami





P.O Box 1982


[email protected]

9th Maryam Temitayo Ahmed





P.O Box 1982


[email protected]

Abstract— With the advancement of technology, we continue

to discover new ways of providing technological solutions in a

variety of fields. Healthcare is one of those fields that can benefit

from technology. With the assistance of technology, the healthcare

field can perform new discoveries, collect more data on symptoms

of diseases, and provide greater support and care to their

patients.[1][2] ADHD (Attention-Deficit/Hyperactivity Disorder)

is a neurodevelopment disorder that is typically identified during

childhood and lasts into adulthood. Lack of focus, inability to sit

still for long periods of time, forgetfulness, impulsivity, time

management difficulties and disorganization, and emotional

dysregulations are some of the early indications of this illness.

Currently, pharmaceutical and nonpharmacological therapies are

used to treat ADHD patients [3]. However, in addition to medical

diagnosis and treatment, it is one of the diseases in which patients

can benefit from technological solutions such as time

management, task management, mindfulness, etc. In this paper,

we propose the requirements and design of a mobile application

called "Flourish," which is aimed at individuals with Attention-

Deficit/Hyperactivity Disorder (ADHD) to help them better

manage their symptoms and understand how they behave. This

application will be divided into sections to assist users in dealing

with the various symptoms experienced by ADHD patients. It

includes time management and organization, mindfulness and

relaxation, a platform for communicating with other ADHD

patients and specialists, habit tracking, and behavioral patterns.

In the future, we intend to develop a fully working real-time

Android application that will be available to assist patients in

managing their symptoms.

Keywords—ADHD, Use Case Diagram, Android OS, Prototypes,

Communication Platform, mobile application, artificial

intelligence, Axure Software.

I. INTRODUCTION

Attention-Deficit/Hyperactivity Disorder (ADHD) is a

neurodevelopment disorder that is typically identified during

childhood and lasts into adulthood [4]. The world health

organization (WHO) prioritizes the ADHD problem because

if individuals with ADHD are not treated early on, the

symptoms will have an impact on learning and other activities

throughout their lives, especially in youngsters [5]. The early

symptoms of this condition include a lack of focus, being

unable to remain still for a long period of time, forgetfulness,

impulsivity, time management difficulties and

disorganization, and emotional dysregulations [6][7]. ADHD

patients can depict these symptoms all at once and diagnosis










2 | P a g e


is done through medical tests that include hearing and vision

examinations. The treatment is usually a combination of both

behavioral therapy and medication [8]. Along with treatment

adherence, patients may still need external assistance, this

had led our team to develop the concept of “Flourish” which

is an application meant to provide better support to ADHD

patients without needing to rely on other people. With the

help of technology, we aim to provide an improved quality of

life to the patients by giving them a tool that helps them with

their daily struggles. There have been a variety of studies and

projects that all have the same goal, helping ADHD patients.

In this paper we explore the different systems that have been

developed for those purposes as well as explaining how

“Flourish” is being developed with multiple features that give

support to ADHD patients and also allows them to be more

independent. The

application will consist of four main sections which are:

organization and management, mindfulness and relaxation, a

platform that connects them with others, and lastly a section

that focuses on habit tracking and behavioral patterns which

is an important part in managing the symptoms of ADHD.

The management and organization section will help in

organizing and prioritizing tasks to make an efficient and

effective timetable. The mindfulness and relaxation section

will help them with their emotional instability, the users will

be able to alleviate their symptoms by going through a

wonderful experience which will help them be more relaxed

and comfortable. ADHD patients need to feel contained, so

Flourish will give them a platform that will enable the

patients to communicate with one another and to also

communicate with specialist in the ADHD field (doctors,

counsellors, coaches...etc.) this can be achieved through

direct messaging between users on the application. The last

section which tracks habits, and any behavioral patterns can

aid in the diagnosis or treatment process.

II. BACKGROUND AND LITERATURE REVIEW

ADHD patients face several challenges in their life, from

time management issues to feeling isolated finding the

community who faces the same problems and so on [9]. In

order to help them and make their life easier we used the

power of technology. Many applications were developed to

help them, but each application was specified to solve one

problem, therefore we decided to build an application that

combines all the separate functionalities in one application,

and we added more ideas to improve the experience of the

users. The goal is to provide a platform that contains all the

resources that can help ADHD patients in an effective

manner. The following is a description of existing real-life

systems that were designed to help ADHD patients, along

with the features and limitations of these systems.

A. Brili Routines - ADHD Habit Tracker

This application meant to help ADHD patients achieving

their daily tasks, support them whenever they feel sad and

overwhelmed, and help them to have and build healthy

routines and habits. The limitation of this application is that

it is not free and there are some privacy issues regarding

accessing the photos [10].

B. ADHD - Cognitive Research

The application is made to collect participations from people

who wants to be involved in an ADHD scientific study. Users

can get their evaluation after finishing the assessment and

download the results. The limitation is that there are no other

features that ADHD patient can enjoy or benefit from in the

application [11].

C. My ADHD

My ADHD application provides diagnostic test to help the

users who are curious to see the results. It provides ARSA 1

for adult users, and SNAP-IV for younger users. The

application has three main sections: articles about ADHD,

techniques, and the diagnostics tests. The limitations are that

the application does not include the references of the articles

that it is showing, along with that these articles are limited,

and it is not informative enough for ADHD patients [12].

D. Brain Focus

The application helps ADHD patients to manage their time

and encourage them to finish their tasks. Users can also group

their tasks together based on certain categories, it also

supports multiple languages, and users can receive

notifications when their work session is done. The limitation

here is that there are some issues with the timer, where users

cannot edit the time after setting it [13].

E. SimpleMind

This application uses diagrams and maps to organize user’s

thoughts and illustrate them with the maps they desire. The

users can draw and create as complex diagrams as they want

and keep track with their ideas and thoughts, but the only

limitation with this application is that it is not free [14].

F. ProductiveHabit Tracker

It is also an application that helps users to build healthy habits

and keep track with the progress using statistics and

encourage them to achieve their goals. It is fully

customizable, and it demonstrates a summary of the user’s

tasks and gives them a reminder when there is a delay in

accomplishing the task. The limitation with this application

is that it is not easy to use, and it might be difficult to navigate

through especially for novice users [15].

G. Due Reminders & Timers

This application has the goal to make sure that all the

commitments and obligations of the users are met and on

time. It keeps reminding users of the tasks that they need to

accomplish and keep track of their progress. It has deferent

features where users can countdown the timer, snooze the

notifications, synchronize their tasks with other devices, and

it has many themes that the user can pick from. The limitation

is that the application is not free, and it is not available for

Android phones users [15].

Finally, the Flourish application is designed with the goal of

providing an easy-to-use platform for ADHD patients while

also combining all the resources that they might need in one

place. The application consists of four main sections:

Behavior Tracking, Mindfulness and Relaxation, Time

Management, and Communication. The features of the

application are as follows: it only works for Android devices,


3 | P a g e


it allows users to have the community of other ADHD

patients and having the ability to communicate with them, it

also helps users with their tasks and organize them in

categories that they desire, and it helps users by providing

them with methods and techniques to regulate their emotions

and have a visualization of their behavioral patterns.

III. PROBLEM STATEMENT

ADHD patients are prone to facing challenges daily, those

challenges can vary from time management and organization

challenges to emotional dysregulation challenges, in addition

to such issues they can feel overwhelmed with the number of

tasks and obligation they need to complete all at once, this

can likely lead to missing important deadlines or big events.

Research shows that people with ADHD face many troubles

regarding emotional dysregulation as they experience several

symptoms such as mood fluctuations, irritability, and temper

tantrums [14]. Due to that kind of distress, they often require

external assistance to be able to stay motivated and no

succumb to their emotions, however that kind of support from

individuals may not always be available to them as they aren’t

always surrounded by people to do so [16]. Therefore, we

have proposed the concept of “Flourish” which can provide

support and availability at all times of the day. The main goal

of the project is to give individuals with ADHD a tool that

allows them to be more independent, gives them the support

they need throughout the day, and to also lessening the

feelings of isolation they may face by providing a platform

where they can communicate with others who experience the

same struggles as them [17].

IV. PROPOSED MODEL STRUCTURE

The proposed model is an artificial intelligence-based

Android application. The goal of a Flourish app is to help

users prioritize their chores based on two factors: deadlines

and difficulty. After that, the user will be given a schedule

with their duties prioritized, allowing them to conveniently

begin and complete tasks. A reward will be given once each

activity is done to serve as motivation. The model structure,

on the other hand, Users may have trouble remembering what

they need to accomplish, so there will be a reminder that will

send an alert when the task approaches its deadline. Other

types of support, such as emotional support, a space to speak

with other ADHD patients or healthcare experts and

specialists, and a component that helps patients track their

habits and uncover their behavioral patterns, will all be

available through our applications.

A. Use case diagram

The use case diagram includes each functional requirement

listed in detail in a table that included the requirement

number, type, name, description, rationale, priority, and

dependencies. Other non-functional requirement included:

Figure 1. system design

Figure 2: Use case diagram


4 | P a g e


performance requirement, safety requirement, usability,

security requirement, and software quality attributes will be

considered while developing the real time application [18].

B. Requirements and Functions.

The following functionalities will be included in the

proposed mobile application.

A. Task management

The user can input the key tasks they need to complete,

as well as their importance level, and these activities will be

organized properly.

B. Mindfulness and Relaxation

This section contains an explanation of the functional needs and their relevance in assisting ADHD patients with emotion regulation. This feature aids the system in recognizing the problem and suggesting several approaches or tactics for resolving it.

C. Communication Platform

This section describes the functional criteria for

improving communication between patients and other users.

Help patients and monitors to communicate directly.

D. Habit Tracking and Behavioral Patterns

This part will store the ADHD patients' behavioral

patterns and follow their activities. This requirement assists

the user in better understanding what they are doing and

how they act.

E. Manage post

The user can update, remove, or add more information to

their thread. Patients can also respond to one another's

threads. Monitors have the ability to respond as well. Users

can report any objectionable discussions or comments. It aids

in the changing of threads so that others may comprehend

what the sufferer is writing. Patients can help each other by

posting to posts, while monitors can write advise and assist

patients by reacting to threads. With reporting, the

application's quality and efficiency will be improved, as well

as user respect.

V. USER CLASSES AND CHARASTRISTICS

The Flourish app provides services to two categories of

users: children (up to the age of 12) and adult patients (from

13 and above). If the user is a child, their parents or caregivers

will be responsible for logging in and using the application's

capabilities, as children are not allowed to use electronics (as

recommended by doctors), and they have nearly the same

functionality as older users. Older users, on the other hand,

can access all the app's features. There are also administrators

who can oversee the system's users.

VI. SYSTEM DESIGN (PROTOTYPE)

To illustate the interfaces of the application a prototype

was created using the Axure software, the main

interfaces are shown in the figures below along with

breif descriptions of their purpose.

3. Task interface

Figure 4 shows the interface for the organization section where

the user is able view their pending tasks and their progress

Figure 4. Task interface

1. Splash screen

The interface below is the first thing that showed once the

user clicks on the application, it shows the logo of the

application and the logo of the University.

Figure 3. Flash screen interface

2. Login interface

The login interface which allows the user to sign in or sign

up if they don’t have an account to access the features.

Figure 3. Login interface


5 | P a g e


6. Behavioral patterns interface

Below interface shows the pattern in a diagram for the

user to make it clearer and helps to track their behavior.

7. Recommended technique interface

Whenever the user feels uncomfortable, through these

interfaces they can handle this feeling. Above interface

displays the result and the application’s recommendation.

It also shows the name of technique and its steps to the

user.

VII. CONCLUSION

The advent of a wide variety of mobile apps that demand

usable interfaces has fueled the rise of mHealth due to the

rapid distribution of smart mobile devices and their immense

potential. This work comprises a requirement description and

prototype for an Android mobile application that can be

created to assist ADHD patients with task and time

management. By introducing mindfulness exercises, it will

also assist patients in moderating their behavior. The

application includes a chat feature that allows users to

communicate with each other.

4. Edit task interface

To edit a task’s name and priority the user can click on the

task from the interface in figure 4 and they will be taken to the

interface in figure 5.

Figure 5. Edit task interface

5. Chat interface and selection of person

The chat section allows the users to interact with one

another via text or voice recordings, they are also shown

a list where they can select who they want to start a chat

with. Figure 6 shows the layout of the interface.

Figure 6. Chat interface

Figure 7. Pattern interface

Figure 8. Technique interface

8. Technique steps interface

In figure 9, the steps of the recommended

techniques are shown to the user, they are able to

move from one step to the other through “next

step” button, and they may also stop the session

if they need to.

Figure 9. Steps interface


6 | P a g e


VIII. FUTURE WORK

As per the planning of the team, in addition to the high-fidelity prototype, a fully functional Android application will be implemented. Some of the features of the application will require the use of speech recognition algorithms to run the application as intended.

REFERENCES

[1] FAROOQUI, M.; GULL, H.; ILYAS, M.; IQBAL, S.;

KHAN, M.A.; KRISHNA, G.; AHMED, M. IMPROVING

MENTAL HEALTHCARE USING A HUMAN CENTERED

INTERNET OF THINGS MODEL AND EMBEDDING

HOMOMORPHIC ENCRYPTION SCHEME FOR CLOUD

SECURITY. J. COMPUT. THEOR. NANOSCI. 2019, 16, 1806–

1812

[2] H. Gull, G. Krishna, M. I. Aldossary and S. Z. Iqbal,

"Severity Prediction of COVID-19 Patients Using

Machine Learning Classification Algorithms: A Case

Study of Small City in Pakistan with Minimal Health

Facility," 2020 IEEE 6th International Conference on

Computer and Communications (ICCC), 2020, pp. 1537-

1541, doi: 10.1109/ICCC51575.2020.9344984.

[3] Powell, L., Parker, J., Robertson, N., & Harpin, V.

(2017). Attention Deficit Hyperactivity Disorder: Is

There an App for That? Suitability Assessment of Apps

for Children and Young People With ADHD. JMIR

mHealth and uHealth, 5(10), e145.

https://doi.org/10.2196/mhealth.7371

[4] Faraone, S. V., Banaschewski, T., Coghill, D.,

Zheng, Y., Biederman, J., Bellgrove, M. A., . . . Wang,

Y. (2021). The World Federation of ADHD International

Consensus Statement: 208 evidence-based conclusions

about the disorder. Neuroscience & Biobehavioral

Reviews. doi:10.1016/j.neubiorev.2021.01.022

[5]J. Mitrpanont, B. Bousai, N. Soonthornchart, K.

Tuanghirunvimon and T. Mitrpanont, "iCare-ADHD: A

Mobile Application Prototype For Early Child Attention

Deficit Hyperactivity Disorder," 2018 Seventh ICT

International Student Project Conference (ICT-ISPC),

2018, pp. 1-4, doi: 10.1109/ICT-ISPC.2018.8523973.

[6] K. George, "Signs and Symptoms of Attention

Deficit Hyperactivity Disorder in Adults -

ActiveBeat", ActiveBeat, 2019. [Online]. Available:

https://www.activebeat.com/your-health/women/8-signs-

and-symptoms-of-attention-deficit-hyperactivity-

disorder-in-adults/.

[7] Wasserstien "Adults with ADHD", in Adult

Attention Deficit Disorder: Brain Mechanisms and Life

Outcomes, 1st ed., L. E., J. Wasserstein and P.

H.Wended, Ed. 2001.

[8] Centers for Disease Control and Prevention. 2021.

What is ADHD?. [online] Available at:

<https://www.cdc.gov/ncbddd/adhd/facts.html#:~:text=

ADHD%20is%20one%20of%20the,)%2C %20

or%20be%20overly%20active.>

[9] M. Brod, E. Schmitt and M. Goodwin, "ADHD

burden of illness in older adults: a life course

perspective", 2011.

[10] A. Leary, "Exploring the Powerful Link Between

ADHD and Addiction", Healthline, 2019. Accessed on:

Sep. 12, 2021. [Online]. Available:

https://www.healthline.com/health/mental-health/adhd-

and-addiction#1

[11] "ADHD - Cognitive Research", Play.google.com,

2020. Accessed on: Sep.12,2021. [Online]. Available:

https://play.google.com/store/apps/details?id=com.cogni

fit.android.adhd&hl=en-GB.

[12] "My ADHD", Play.google.com, 2021. Accessed on:

Sep.12,2021. [Online].Available:

https://play.google.com/store/apps/details?id=tk.tdah&hl

=en-.

[13] T. Angel, "Everything You Need to Know About

ADHD", Healthline, 2021. Accessed on: Oct.20 ,2021.

[Online]. Available:

https://www.healthline.com/health/adhd

[15] A. Doyle, "The 11 Best ADHD Apps for

2022", Healthline, 2021. .

[16] "Index to Journal of Attention Disorders", Journal

of Attention Disorders, vol. 9, no. 4, pp. 669-673, 2006.

Available: 10.1177/108705470600900414 [Accessed 11

September 2021].

[17] "Erratum", Child and Adolescent Psychiatric

Clinics of North America, vol. 9, no. 3, pp. Pages 481-

498, 2016. Available:

https://www.sciencedirect.com/science/article/abs/pii/S1

056499318301020. [Accessed 13 September 2021].

[18] [12]W. Shen and S. Liu, International Conference

on Formal Engineering Methods, 1st ed. Piscataway,

N.J.: IEEE, 2012, pp. pp 68–85.

IEEE conference templates contain guidance text for

composing and formatting conference papers. Please

ensure that all template text is removed from your

conference paper prior to submission to the

conference. Failure to remove template text from

your paper may result in your paper not being

published.

https://doi.org/10.2196/mhealth.7371


1 | P a g e Published In: 7th International Conference on Data Science and Machine Learning Applications (CDMA) IEEE DOI: 10.1109/CDMA54072.2022.00024

Machine Learning Based Preemptive Diagnosis of Lung Cancer Using Clinical

Data

Sunday O. Olatunji1, Aisha Alansari2, Heba Alkhorasani3, Meelaf Alsubaii4, Rasha Sakloua5, Reem Alzah-

rani6, Yasmeen Alsaleem7, Reem Alassaf8, Mehwash Farooqui9, Mohammed Imran Basheer Ahmed10

College of Computer Science and Information Technology, Imam Abdulrahman Bin Faisal University, P.O. Box

1982, Dammam 31441, Saudi Arabia

[email protected] , [email protected], [email protected], [email protected], [email protected], [email protected], [email protected], [email protected],

[email protected], [email protected]

Abstract— Lung cancer is a malignant disease that im-

poses serious complications restricting patients from

performing daily tasks in the early stages and eventu-

ally cause their death. The prevalence of this disease

has been highlighted by numerous statistics worldwide.

The preemptive diagnosis of individuals with lung can-

cer can enhance chances of prevention and treatment.

Therefore, the purpose of this study is to predict lung

cancer preemptively utilizing simple clinical and demo-

graphical features obtained from the “data world”

website. The experiment was conducted using Support

Vector Machine (SVM), K-Nearest Neighbor (K-NN),

and Logistic Regression (LR) classifiers. To improve

models’ accuracy, SMOTETomek was employed along

with GridsearchCV to tune hyperparameters. The Re-

cursive Feature Elimination method was also utilized to

find the best feature subset. Results indicated that SVM

achieved the best performance with 98.33% recall,

96.72% precision, and an accuracy of 97.27% using 15

attributes.

Keywords— Support Vector Machine, Logistic Regres-

sion, K-Nearest Neighbor, Machine Learning, Lung

Cancer, Preemptive Diagnosis.

I. INTRODUCTION 

Lung cancer is a type of cancer where the cells in one

or both lungs grow uncontrollably, forming tumors [1].

The cancer cells spread to other organs once it reaches

stage 4, the advanced stage, leading to death [2]. How-

ever, if lung cancer is discovered in its early stage, the

treatment is easy, which can successfully keep the pa-

tient from dying [3]. Unfortunately, the early diagnosis

of the disease is challenging since it can stay silent with-

out any symptoms for over 20 years before it reaches its

advanced stage and becomes lethal [1]. Thus, it is con-

sidered one of the deadliest cancers among other types

of cancer. In 2020, the Global Cancer Observatory

(GCO) reported 2,206,771 new cases along with

1,796,144 deaths internationally, ranking it as the sec-

ond and first, respectively [4]. It is expected that ap-

proximately 150 million people will pass away from

lung cancer in the 21st century. As a result, preventing

and improving the state of patients with lung cancer is a

top international priority [5]. This study aims to take ad-

vantage of new technologies by employing machine

learning techniques in the pre-emptive diagnosis of lung

cancer.

Machine learning is one of the widely utilized Artificial

Intelligence techniques, which consists of various algo-

rithms that are trained and tested based on a dataset to

generate predictions [6]. It offers a variety of tech-

niques, methods, and tools that are beneficial for solv-

ing challenging problems in the medical domain, in-

cluding the early diagnosis of a disease. Machine learn-

ing was ultimately deemed to be successful in improv-

ing the overall quality and efficiency of medical care.

Among the several supervised machine learning tech-

niques, three algorithms shown to be useful in the med-

ical field were employed in this study, namely, Support

Vector Machine (SVM), K-Nearest Neighbor (K-NN),

and Logistic Regression (LR). SVM is the most used al-

gorithm in disease prediction due to its superior predic-

tive accuracy, whereas K-NN is considered a simple al-

gorithm with fast instances classification. On the other

hand, LR can be easily implemented and updated

[7]. The purpose of this study is to predict Lung cancer

in its early stage using simple clinical data. The empiri-

cal results of this study showed that SVM and LR

achieved the highest accuracy rates of 97.27% with 15

features. However, from a medical perspective, SVM

was proven to be the best classifier since it granted

the highest recall rate of 98.33%.  

In this study, the remaining sections are organized in the

following order. The second section covers the litera-

ture review of related works. The third section contains

a description of the three utilized machine learning

about:blank

about:blank

about:blank

about:blank

about:blank

about:blank

about:blank

about:blank

about:blank



algorithms: SVM, K-NN, and LR. The fourth section il-

lustrates the empirical study, including dataset descrip-

tion, experimental setup, performance measures, and

optimization strategy. The fifth section demonstrates

and discusses the study results, while the last section

contains the conclusion and future work recommenda-

tions for this study.

II. RELATED WORK 

The authors in [8] aimed to construct an early Lung can-

cer detection tool to assist general physicians in diag-

nosing cancer 3-6 months before ordinary diagnosis.

The used dataset established by the Australian Govern-

ment Department of Health contained millions of diag-

nosis records, pathology test results, and others. In ad-

dition to ensemble techniques, LightGBM, XGBoost,

Decision Tree (DT), and AdaBoost classifiers were uti-

lized to compose the model. The results indicated that

the ensemble model outperformed the other models

considering the true-positive rate (45%) and the false-

negative rate (35%). 

Furthermore, the authors in [9] developed a system to

predict Lung cancer based on multi-criteria decisions.

They gathered the dataset from a web-based survey of

276 individuals with eight attributes. The authors used

analytic hierarchy process (AHP) to assign weights used

by the Artificial Neural Network (ANN) classifier. The

experimental result and discussion showed that the pro-

posed model achieved an accuracy of 80.7%, specificity

of 75.3%, sensitivity of 89.9%, and an F1 score of

86.4%.  

Moreover, work in [10] utilized two online datasets to

predict lung cancer using various machine learning al-

gorithms. The first dataset was retrieved from the UCI

website comprising 56 predictors to pre-diagnose three

types of Lung cancer. It was concluded that the Deci-

sion Tree (DT) algorithm achieved the highest accuracy

of 96.9%. Afterward, the authors used the second da-

taset, acquired from the Data-World website, containing

23 clinical features with a 3-stage target class. The re-

sults demonstrated that SVM exceeded all other classi-

fiers with an accuracy of 99.2%.  

The study [11] used the same UCI dataset with deep

learning algorithms to early classify Lung cancer. The

recurrent neural network (RNN) and BiLSTM were ap-

plied, achieving the average specificity, sensitivity, pre-

cision, recall, and f-score values were 0.96, 0.98, 0.97,

0.96, and 0.96, respectively. 

Meanwhile, the same Data-World website dataset was

used by researchers in [12]. They implemented a hy-

bridization approach between the Genetic Algorithm

(GA), used for feature selection, and the K-Nearest

Neighbor (K-NN), used for classification, to develop a

Lung cancer prognosis model. As a result, the K-NN al-

gorithm revealed a 100% accuracy with six optimal fea-

tures captured by GA. The same dataset was used in

[13], where the authors proposed the analysis of Lung

cancer symptoms for different age groups in the early

stage using machine learning techniques. The result

showed that DT, RF, and XGBoost reached the highest

accuracy of 100% in the Youth and Working group and

93% in the Elderly. 

The review of literature related to this study showed no

studies that predicted Lung cancer pre-emptively with

high accuracy using simple clinical and demographical

data. Additionally, to the best of our knowledge, no pre-

vious studies utilized the same dataset used in this

study. Most of the studies attaining high accuracy fo-

cused on predicting the severity rate of the disease.

Therefore, this study explores the effect of employing

simple clinical and demographical data in the pre-emp-

tive diagnosis of Lung cancer using machine learning

techniques to reduce the possible hazards of late detec-

tion of Lung cancer.  

III. DESCRIPTION OF PROPOSED TECH-

NIQUES 

A. Support Vector Machine (SVM) 

Support Vector Machine (SVM) is a popular machine

learning algorithm that researchers successively uti-

lize for its ability to form robust results regardless of

noisy and scarce data [14]. It is a statistical-based super-

vised algorithm, mainly employed in binary classifica-

tion problems, where it operates by constructing a

boundary that divides the training data into two sepa-

rated classes [15]. This boundary is known as the hyper-

plane, which is a subspace with dimension p-1. In order

to find the optimal hyperplane, the margin, which is the

distance between the line and the nearest points to the

line, called support vectors, is calculated [16]. Further

details could be found in [17][18].  

B. Logistic Regression (LR) 

Logistic Regression (LR) is a machine learning algo-

rithm based on supervised learning and statistical anal-

ysis. It is extensively used for binary classification tasks

with discrete categorical target classes [19]. In 1958, LR

was developed by David Cox and named after the lo-

gistic function, which is the core of its process [20]. It

applies the sigmoid function to a linear combination of

features to restrict them between 0 and 1. Afterward, it

compares the result with the threshold value, which is



equal to 0.5 by default. The result is assigned to the pos-

itive class if it is greater than the threshold

value, whereas it is assigned to the negative class if it is

less than the threshold value [19]. Further details could

be found in [21][22].   

C. K-Nearest Neighbor (K-NN) 

K-Nearest Neighbor (K-NN) is one of the simplest non-

parametric machine learning algorithms that fall below

the supervised learning technique [23]. K-NN is consid-

ered a lazy learner as it does not learn from training data,

but it uses the entire dataset to classify the data points

without building a training model  [24]. It categorizes a

new data point by following how its neighbors are clas-

sified and predicting its value based on how closely it

supplements the training set points using a similarity

measure [25].  Further details could be found in

[26][27].   

IV. EMPIRICAL STUDIES

A. Description of Dataset

The dataset used for the preemptive prediction of lung

cancer was obtained from the “data world” website,

consisting of clinical and demographical features [28].

The dataset included 16 attributes and 309 instances.

270 instances were lung cancer patients, whereas 39 in-

stances were non-lung cancer patients. After applying

the SMOTETomek sampling technique [29], the data

consisted of 182 positive patients and 182 negative pa-

tients. Table 1 outlines each attribute with its type and

corresponding values.

Table 1 Attribute Description

Attribute  Type  Value 

Gender  Nominal  Male: M, Female: F 

Age  Numeric  Age of the patient: No range 

Smoking  Numeric  Yes: 2, No:1 

Yellow Fingers  Numeric  Yes: 2, No:1 

Anxiety  Numeric  Yes: 2, No:1 

Peer_Pressure  Numeric  Yes: 2, No:1 

Chronic Disease  Numeric  Yes: 2, No:1 

Fatigue  Numeric  Yes: 2, No:1 

Allergy  Numeric  Yes: 2, No:1 

Wheezing  Numeric  Yes: 2, No:1 

Alcohol  Numeric  Yes: 2, No:1 

Coughing  Numeric  Yes: 2, No:1 

Shortness of Breath  Numeric  Yes: 2, No:1 

Swallowing Diffi-

culty  Numeric  Yes: 2, No:1 

Chest Pain  Numeric  Yes: 2, No:1 

Lung_Can-

cer (Class)  Nominal 

Lung cancer: yes, no lung

cancer: no 

B. Statistical Analysis of the Dataset

Table 2 shows the dataset’s statistical analysis, includ-

ing the mean, median, standard deviation, and the cor-

relation between each attribute and the target attribute.

The low standard deviation values demonstrate that the

data are close to the mean, indicating no outliers in the

data. Besides, the similarity between the mean and me-

dian values indicates a symmetric distribution of the

data.

Table 2 Statistical Analysis of the Dataset

Attribute  Mean  Median   Standard

Deviation  

Correlation

Coefficient  

Gender - - - 0.067254  

Age  62.673139   62.0   8.210301   0.089465  

Smoking   1.563107   2.0   0.496806   0.058179  

Yellow

Fingers  1.569579   2.0   0.495938   0.181339  

Anxiety   1.498382   1.0   0.500808   0.144947  

Peer_Pressure  1.501618   2.0   0.500808   0.186388  

Chronic

Disease   1.504854   2.0   0.500787   0.110891  

Fatigue  1.673139   2.0   0.469827   0.150673  

Allergy   1.556634   2.0   0.497588   0.327766  

Wheezing   1.556634   2.0   0.497588   0.249300  

Alcohol   1.556634   2.0   0.497588   0.288533  

Coughing   1.579288   2.0   0.494474   0.248570  

Shortness of

Breath  1.640777   2.0   0.480551   0.060738  

Swallowing

Difficulty   1.469256   1.0   0.499863   0.259730  

Chest Pain   1.556634   2.0   0.497588   0.190451  

 

C. Experimental Setup

The experiment was conducted to build a prediction

model that pre-emptively predicts the presence of lung

cancer using simple clinical and demographical fea-

tures. Python programming language was employed to

conduct the experiment, where several libraries were

utilized. The Imblearn library was utilized to balance

the dataset using the SMOTETomek technique with

random state value 139, which performs over-sampling

using the Synthetic Minority Oversampling Technique

(SMOTE) followed by Tomek links that clean the da-

taset [29]. The Sklearn library was utilized to split the

dataset into 70% for training and 30% for testing. Addi-

tionally, it was used to perform the GridsearchCV tech-

nique to obtain the best hyper-parameters using 10-folds

cross-validation. The best obtained hyper-parameters

were then utilized to build three models from the same

library, namely, Support Vector Machine (SVM), Lo-

gistic Regression (LR), and K-Nearest Neighbor (K-

NN) using the 70:30 direct partition technique with the

random state value 0. Subsequently, the Recursive



Feature Elimination method was examined to find the

best feature subset yielding the highest performance by

relying on the correlation coefficient calculated by the

Pandas library.

D. Performance Measures

In this study, three performance measures, including

precision, recall, and accuracy, were utilized to evaluate

the performance effectiveness. The precision calculates

the lowest limit of incorrectly classified as lung cancer

(FP), whereas recall aims to find the lowest limit of in-

correctly classified as no lung cancer (FN). On the other

hand, the accuracy evaluates the number of correct pre-

dictions. The equations below show the formulas of

each performance metric.

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃

𝑇𝑃+𝐹𝑃 (1)

𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃

𝑇𝑃+𝐹𝑁 (2)

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃+𝑇𝑁

𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑖𝑛𝑠𝑡𝑎𝑛𝑐𝑒𝑠 (3)

E. Optimization Strategy

To achieve the best solution among all possible solu-

tions and to solve classification problems optimally, op-

timization tools must be applied. In this study, the grid

search was used to find the optimal hyperparameters

that produce the highest accuracy. Grid search operates

by defining a search space set as a grid of specified hy-

perparameters with a range of values, then tries all pos-

sible combinations to return the hyper-parameters

that achieve the best performance. 10-folds cross-vali-

dation was used for validation in the optimization strat-

egy. 

 

For the SVM, the searched hyperparameters were cost,

kernel, and gamma. The cost values included the values

(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, and 30). Moreover,

the list of kernel types included (RBF, sigmoid, and lin-

ear). Besides, the gamma values included (1, 0.1, 0.01,

0.001, and 0.0001).

The hyperparameters searched for K-NN include

n_neighbors and metric parameters. The n_neighbors

values searched were (5, 7, 9, 11, 13, 15, 17, 19, 21, 23,

25, 27, 29, 31, 33, 35, 37, and 39). Furthermore, the met-

rics included (Minkowski, Euclidean, and Manhattan).

The Logistic Regression hyperparameters searched

were cost, penalty, and slover. The cost values include

100, 10, 1.0, 0.1, and 0.01. Moreover, the solver values

included (Newton-cg, Lbfgs, Liblinear, Sag, and Saga).

Besides, the penalty values included (None, L1, L2, and

Elasticnet). Table 4 shows the chosen optimal hyper-pa-

rameters for each classifier.  

 

Table 4 The Hyper-parameters for Each Classifier

Classi-

fier 

Hyperparame-

ter 

Values Without

Sampling 

Values With

Sampling

SVM 

Cost  7  15 

Gamma  1  0.001 

Kernel  Linear  RBF 

K-NN  n_neighbors  11  5 

metric  Manhattan  Manhattan 

 LR 

Cost  100  0.1 

Penalty  12  12 

Solver  Liblinear  Newton-cg 

V. RESULTS AND DISCUSSION

The optimal hyperparameters obtained from the

GridsearchCV technique were used to build the models

using the 70:30 direct partition technique. The achieved

training and testing accuracy before and after applying

the SMOTETomek using the optimal hyper-parameters

are outlined in Table 5.

  Table 5 Classifiers Training and Testing Accuracy Using the Opti-

mal Hyper-parameters

Classi-

fier 

Sampling

Technique 

Training

Accuracy 

Testing

Accuracy 

SVM 

Without Sam-

pling  93.98%  93.55% 

With Sampling 97.24%  97.27% 

K-NN 

Without Sam-

pling  86.57%  92.47% 


LR 

Without Sam-

pling  93.52%  94.62% 


 

It is observed that the difference between the training

and testing accuracy is low before and after applying the

sampling technique, which indicates that the models

neither suffered from overfitting nor underfitting. Addi-

tionally, it is noted that the SMOTETomek technique

boosted the testing accuracy of both SVM and LR by

3.72% and 2.62%, respectively. However, it did not

show a remarkable effect on the performance of K-NN.

The highest results were achieved by SVM and LR,

reaching an accuracy of 97.27% after applying the

SMOTETomek sampling technique. In the upcoming

sections, further analysis will be done on the data after

applying feature selection on the SMOTETomek gener-

ated data.

A. Results of Investigating the Effect of Feature Se-

lection on the Dataset

Using the correlation coefficient presented in Table 3,

several feature subsets were selected to compare their

effect on the performance using the best hyper-parame-

ters. The correlation coefficient helps rank the attributes

in descending order based on the correlation value



between each attribute with the target class. The Recur-

sive Feature Elimination technique was carried out to

obtain the most appropriate feature subset that provides

the highest results. Table 6 outlines the average accu-

racy of different feature subsets. 

Table 6 Average Accuracy for Different Feature Subsets

Features  SVM  K-NN  LR  Average 

Using 15 Features  97.27%  92.73%  97.27%  95.76% 

Using 8 Features  94.55%  90.91%  94.55%  93.33% 

Using 4 Features  91.82%  94.55%  91.82%  92.73% 

Using 2 Features  81.82%  90.00%  90.00%  87.27% 

Using 1 Features  81.82%  81.82%  81.82%  81.82% 

 

The empirical results indicated that the Recursive Fea-

ture Elimination Technique did not improve the accu-

racy of the models, where the highest average accuracy

of 95.76% was attained using 15 features (full features).

This might be caused by the low correlation values be-

tween the attributes and the target class. Therefore, all

the features present in the dataset will be utilized for

building the final models with the best obtained hyper-

parameters.  

B. Further Discussion of the Results

The classification performance on the SMOTETomek

generated data using the optimal parameters and full

features for SVM, K-NN, and LR are illustrated in table

7. 

Table 7 Classification Performance of the Final Models

Classifier  Accuracy  Precision  Recall 

SVM  97.27%  96.72%  98.33% 

K-NN  92.73%  98.15%  88.33% 

LR  97.27%  98.31%  96.67% 

The results indicated that both SVM and LR outper-

formed K-NN with an accuracy of 97.27%. However,

further evaluation measures were executed by analyzing

recall and precision values. SVM produced the highest

recall rate among all other algorithms with 98.33%,

while LR attained the highest precision rate with

98.31%. To investigate the consistency between pre-

dicted and actual findings, confusion metrics were used.

Tables 8, 9, and 10 illustrate the confusion matrices that

characterized the prediction results by the proposed

models.

  Table 8 SVM Confusion Matrix

SVM  Predicted 

Lung Cancer  No Lung Cancer 

Actual  Lung Cancer   59 (TP)  1 (FN) 

No Lung Can-

cer  2 (FP)  48 (TN) 

  Table 9 K-NN Confusion Matrix

K-NN  Predicted 

Lung Cancer No Lung Cancer

Actual 

Lung Cancer  53 (TP)  7 (FN) 

No Lung Can-

cer  1 (FP)  49 (TN) 

Table 10 LR Confusion Matrix

LR  Predicted 

Lung Cancer  No Lung Cancer 

Actual 

Lung Cancer  58 (TP)  2 (FN) 

No Lung Can-

cer  1 (FP)  49 (TN) 

 

Detecting every potential indication of lung cancer ill-

ness will reduce the error rate and provide patients with

prevention chances. Consequently, to avoid complica-

tions that could arise from misdiagnosing the presence

of a disease, we meticulously examined the number of

false-negative. Hence, the SVM is considered the best

classifier among the investigated ones as it yielded the

lowest FN value of 1, followed by LR with 2 FN, while

K-NN suffered from excessive FN of 7.

VI. CONCLUSION AND RECOMMANDATION

Lung cancer is one of the most dangerous diseases that

are difficult to treat after it spreads and reaches a severe

stage. In this paper, pre-emptive prediction models of

lung cancer were developed using three machine learn-

ing techniques, namely, Support Vector Machine

(SVM), K-Nearest Neighbor (K-NN), and Logistic Re-

gression (LR) using an open-source dataset with clinical

and demographical data. The GridsearchCV technique

with 10-folds cross-validation was utilized to obtain the

best hyper-parameters of each algorithm. Additionally,

the Recursive Feature Elimination technique was inves-

tigated to find the best feature subset. The empirical re-

sults showed that SVM and LR classifiers achieved the

highest accuracy of 97.27% with full features. How-

ever, SVM outperformed LR since it granted the highest

recall rate of 98.33%.

The utilized dataset has a limited number of in-

stances and a weak correlation between the attributes

and the target class. Increasing the number of in-

stances and exploring more correlated attributes with

the target class would contribute to enhancing the re-

sults. Therefore, future work can be done by re-using

the proposed technique with a large dataset and more

features such as Hoarseness, Nail clubbing, and Loss of



appetite/weight. Furthermore, further studies can be

conducted by employing ensemble techniques and ex-

amining other feature selection methods to enhance the

performance with the most crucial features. Addition-

ally, the promising empirical result motivated us to ex-

pand the work by investigating a Saudi dataset to con-

tribute to Saudi Arabia’s vision 2030 in the objectives

of the Transformation program in the future.

REFERENCES

[1] “Chronic Disease - Lung Cancer.”

https://www.moh.gov.sa/en/awarenessplatefor

m/ChronicDisease/Pages/LungCancer.aspx

(accessed Oct. 21, 2021).

[2] P. Zhang et al., “Nanotechnology-enhanced

immunotherapy for metastatic cancer,” Innov.,

p. 100174, Oct. 2021, doi:

10.1016/J.XINN.2021.100174.

[3] M. Santarpia et al., “Liquid biopsy for lung

cancer early detection,” J. Thorac. Dis., vol. 10,

no. 7, pp. S882–S897, Apr. 2018, doi:

10.21037/JTD.2018.03.81.

[4] “World,” Accessed: Nov. 19, 2021. [Online].

Available:

https://gco.iarc.fr/today/data/factsheets/populati

ons/900-world-fact-sheets.pdf.

[5] A. R. Jazieh, G. Algwaiz, S. M. Alshehri, and K.

Alkattan, “Lung Cancer in Saudi Arabia,” J.

Thorac. Oncol., vol. 14, no. 6, pp. 957–962, Jun.

2019, doi: 10.1016/J.JTHO.2019.01.023.

[6] J. A. Nichols, H. W. Herbert Chan, and M. A. B.

Baker, “Machine learning: applications of

artificial intelligence to imaging and diagnosis,”

Biophys. Rev., vol. 11, no. 1, p. 111, Feb. 2019,

doi: 10.1007/S12551-018-0449-9.

[7] S. Uddin, A. Khan, M. E. Hossain, and M. A.

Moni, “Comparing different supervised machine

learning algorithms for disease prediction,”

BMC Med. Inform. Decis. Mak., vol. 19, no. 1,

pp. 1–16, Dec. 2019, doi: 10.1186/S12911-019-

1004-8/FIGURES/12.

[8] “Artificial Intelligence in Medicine: 19th

International Conference on ... - Allan Tucker -

Google Books.” .

[9] B. Al-Bander, Y. A. Fadil, and H. Mahdi,

“Multi-Criteria Decision Support System for

Lung Cancer Prediction,” IOP Conf. Ser. Mater.

Sci. Eng., vol. 1076, no. 1, p. 012036, Feb. 2021,

doi: 10.1088/1757-899X/1076/1/012036.

[10] P. R. Radhika, R. A. S. Nair, and G. Veena, “A

Comparative Study of Lung Cancer Detection

using Machine Learning Algorithms,” Proc.

2019 3rd IEEE Int. Conf. Electr. Comput.

Commun. Technol. ICECCT 2019, Feb. 2019,

doi: 10.1109/ICECCT.2019.8869001.

[11] I. Publication, “DEEP LEARNING BASED

BiLSTM ARCHITECTURE FOR LUNG

CANCER CLASSIFICATION,” doi:

10.34218/IJARET.12.1.2020.045.

[12] N. Maleki, Y. Zeinali, and S. T. A. Niaki, “A k-

NN method for lung cancer prognosis with the

use of a genetic algorithm for feature selection,”

Expert Syst. Appl., vol. 164, p. 113981, Feb.

2021, doi: 10.1016/J.ESWA.2020.113981.

[13] A. Bankar, K. Padamwar, and A. Jahagirdar,

“Symptom analysis using a machine learning

approach for early stage lung cancer,” Proc. 3rd

Int. Conf. Intell. Sustain. Syst. ICISS 2020, pp.

246–250, 2020, doi:

10.1109/ICISS49785.2020.9315904.

[14] T. S. Furey, N. Cristianini, N. Duffy, D. W.

Bednarski, M. Schummer, and D. Haussler,

“Support vector machine classification and

validation of cancer tissue samples using

microarray expression data,” Bioinformatics,

vol. 16, no. 10, pp. 906–914, 2000, doi:

10.1093/bioinformatics/16.10.906.

[15] V. Jakkula, “Tutorial on Support Vector

Machine (SVM),” Accessed: Nov. 19, 2021.

[Online]. Available:

https://course.ccs.neu.edu/cs5100f11/resources/

jakkula.pdf.

[16] Y. Lei, “Individual intelligent method-based

fault diagnosis,” Intell. Fault Diagnosis Remain.

Useful Life Predict. Rotating Mach., pp. 67–174,

Jan. 2017, doi: 10.1016/B978-0-12-811534-

3.00003-2.

[17] J. M. Moguerza and A. Muñoz, “Support Vector

Machines with Applications,” Stat. Sci., vol. 21,

no. 3, pp. 322–336, 2006, doi:

10.1214/088342306000000493.

[18] C. Cortes and V. Vapnik, “Support-vector

networks,” Mach. Learn. 1995 203, vol. 20, no.

3, pp. 273–297, Sep. 1995, doi:

10.1007/BF00994018.

[19] H. Belyadi and A. Haghighat, “Supervised

learning,” Mach. Learn. Guid. Oil Gas Using

Python, pp. 169–295, Jan. 2021, doi:

10.1016/B978-0-12-821929-4.00004-4.

[20] A. Bartosik and H. Whittingham, “Evaluating

safety and toxicity,” Era Artif. Intell. Mach.

Learn. Data Sci. Pharm. Ind., pp. 119–137, Jan.

2021, doi: 10.1016/B978-0-12-820045-

2.00008-8.

[21] R. O. Sinnott, H. Duan, and Y. Sun, “A Case

Study in Big Data Analytics: Exploring Twitter

Sentiment Analysis and the Weather,” Big Data

Princ. Paradig., pp. 357–388, Jan. 2016, doi:

10.1016/B978-0-12-805394-2.00015-5.

[22] E. B. Seufert, “Quantitative Methods for Product

Management,” Free. Econ., pp. 47–82, Jan.



2014, doi: 10.1016/B978-0-12-416690-

5.00003-8.

[23] P. Nadkarni, “Core Technologies: Data Mining

and ‘Big Data,’” Clin. Res. Comput., pp. 187–

204, Jan. 2016, doi: 10.1016/B978-0-12-

803130-8.00010-5.

[24] R. A. Alassaf et al., “Preemptive Diagnosis of

Chronic Kidney Disease Using Machine

Learning Techniques,” Proc. 2018 13th Int.

Conf. Innov. Inf. Technol. IIT 2018, pp. 99–104,

Jan. 2019, doi:

10.1109/INNOVATIONS.2018.8606040.

[25] T. M. D. Ebbels, “Non-linear Methods for the

Analysis of Metabolic Profiles,” Handb.

Metabonomics Metabolomics, pp. 201–226, Jan.

2007, doi: 10.1016/B978-044452841-4/50008-

4.

[26] D. Chanal, N. Yousfi Steiner, R. Petrone, D.

Chamagne, and M.-C. Péra, “Online Diagnosis

of PEM Fuel Cell by Fuzzy C-Means

Clustering,” Ref. Modul. Earth Syst. Environ.

Sci., Jan. 2021, doi: 10.1016/B978-0-12-

819723-3.00099-8.

[27] J. S. Richman, “Multivariate Neighborhood

Sample Entropy: A Method for Data Reduction

and Prediction of Complex Data,” Methods

Enzymol., vol. 487, no. C, pp. 397–408, Jan.

2011, doi: 10.1016/B978-0-12-381270-

4.00013-5.

[28] “survey lung cancer - dataset by sta427ceyin |

data.world.”

https://data.world/sta427ceyin/survey-lung-

cancer (accessed Oct. 29, 2021).

[29] “SMOTETomek — Version 0.9.0.dev0.”

https://imbalanced-

learn.org/dev/references/generated/imblearn.co

mbine.SMOTETomek.html (accessed Nov. 19,

2021).


1 | P a g e

Published In: IEEE IAS Global Conference on Emerging Technologies (GlobConET-2022)

Leen: Web-based Platform for Pet Adoption

Reema Alsuwailem

dept. Computer Information

Systems

Imam Abdulrahmaan Bin Faisal

University


[email protected]

Reema Almobarak


Systems


University


[email protected]

Razan Aboali


Systems


University


[email protected]

Safa Alrubaiea


Systems


University


[email protected]

Jazwa Aldossary


Systems


University


[email protected]

Abstract—This paper concerns pet adoption by building a web-

based platform that supports the idea of using technology for the

pet adoption process in Saudi Arabia, the eastern province in

specific. The difficulty of the adoption process and putting pets up

for adoption is a real problem in our society. In this regard, we put

forward the idea of “Leen” to provide easy and quick services for

this process and make it accessible to all interested people. Our

platform offers many services such as adoption, pet care, donation,

Etc. However, the main point in the “Leen” platform is that all

services provided are free with no fees. A “Leen” platform user

can offer a pet for adoption to find a home with another user

from “Leen”. Also, a user can look for pet care clinics at their

nearest location in the region. Furthermore, a user can directly

donate to trusted adoption associations in Saudi

Arabia. “Leen” platform was built to provide the mentioned

services and more. Eventually, having the platform within reach

of users will provide all the services faster and easier than usual.

Keywords—platform, web-based

I. INTRODUCTION

Due to the remarkable awareness in our society towards dealing with and caring for pets, along with the spread of the “pet adoption” concept in recent years. From this point, we got inspired and came up with our idea to build a web-based platform for all people interested in this field. Our platform aims to act as a midpoint between people who want to offer their pets adoption and those willing to adopt. Therefore, the adoption process will be easier and faster. Besides, providing enjoyment for people who love pets to communicate with others who have the same interests to share their knowledge and personal experiences. All the provided services will be for free. Promoting the principle of free pet adoption without any fees has a positive long-term impact on society. Based on a survey conducted by researchers at the University of Florida in 2011 on (1,928) pet adopters, which aimed to study the impact of free

adoptions on society. As a result of the study, it was found that adoptions which do not require any fees are successful and promoting free adoption may raise the adoption rate without compromising the animal’s life quality, as most users reported they still keep pets they adopted, which were 93% dogs, and 95% cats [1].

II. PROBLEM STATEMENT

Nowadays, we are encountering an issue where people have misleading knowledge about owning a pet, thinking that buying a pet from social media sites is more convenient than adopting one because there is no trusted platform provides the exact requirement they want [2]. However, it is considered a problem due to the potential risks a process carries to the users’ privacy; it may be from fraudulent sellers’ accounts. Thus, the incidence of fraud and deception of users increases significantly; moreover, most social media sites focus solely on money and treat pets as a tool to gain money. In order to solve this problem building a reliable integrated website, which acts as a third-party platform that aims to establish the principle of adopting pets instead of the process of buying one, facilitating user experience, as well as simplifying the process of communication between users to be direct while preserving their privacy and providing completely free services.

III. LITERATURE REVIEW

In the literature reviews below, there have been some studies by researchers proposing implementing technology with pets. This section will discuss how our website technology differs from other authors’ related technology regarding similarities and dissimilarities in pet adoption.







2 | P a g e


A. Effect of Visitor Perspective on Adoption Decisions at One

Animal Shelter

One of the methods in adopting a pet is a walk-in shelter to look for an animal. According to a study conducted at one urban animal shelter, adopters can interact with the animal and see if the animal is friendly, energetic, affectionate, and see physical characteristics and looks. In addition, the adopter can also know if the animal is not interactive, friendly, energetic; therefore, it helps the adopter pick an animal according to these characteristics while visiting a shelter. The researcher of this article found that many visitors did not leave with a pet; moreover, some visitors had an intention to adopt a pet but ended up not adopting one. This article categorizes visitors into two categories; the first is just visitors who are not interested in adopting a pet; according to the shelter staff, these visitors waste their time, energy, and resources. The second category is the browser, and they are people who visit the shelter for months and weeks intending to adopt a pet in the future. The writers of this article suggest the shelter work on educating all visitors about animal care and welfare. Moreover, educating those who are not prepared for a new pet [3].

B. How social media helps shelter animals out of the Shadow

Ariel wrote an article about how social media helps animals to be adopted. Social networking sites motivate people to adopt animals. For example, putting a picture of an animal for adoption on a social networking site such as Facebook, Twitter, and Instagram, and sharing it with friends or liking the picture and republishing it helps in animal adoption. Nevertheless, in social media, not all animals get their share of people’s likes [4].

C. A Review of Techniques for Image Classification to

Enhance Online Animal Adoption Speed

According to Pradeepa, animals can be adopted faster using the internet and technology. In addition, developing a computerized application that uses the sheltered animal picture gives it a score to help predict how fast the animal will be adopted; thus, this will help guide the shelter’s animal adoption speed process when posting a picture of the animal. However, the adoption speed cannot be controlled; some animals are not adopted due to having blurry pictures or not being wanted by anyone for adoption; thus, shelters will be overcrowded with animals [5].

D. The Impact of Adopting a Pet in the Perception of Physical

and Emotional Wellbeing

One of the most common reasons people give for possessing pets is the fact that they provide unconditional companionship and a sense of care and protection. However, this study looked at the impact of pets in a therapeutic context and primarily focused on the benefits of keeping pets. It has been found that owning dogs increases their owners’ physical activity and become less likely to have diseases such as obesity. On the other hand, it shows that people who have a pet are more prone to allergies and asthma. Furthermore, the study found that the humanization of pets was key to the emotional impact that adopters perceive, and those who tend to humanize their pets develop an empathetic relationship with them [6].

E. Attitudes and Perceptions Regarding Pet Adoption

This paper discussed the current trends in pet overpopulation and compared findings regarding purchasing from for-profit sources versus adoption from shelters. A survey was sent to registered dog owners in Albany and Rensselaer counties. The findings illustrate that people who are looking for a specific breed and have misperceptions of purebred dogs’ costs tend to go to pet stores and breeders primarily to purchase dogs. As a result, they believe it cannot be satisfied by adopting a shelter dog. However, this study contains two problems. First, responses biases. Second, the lack of comprehensiveness of the study results due to it reflects the respondents’ attitudes in specific regions and does not represent an entire country. In addition, consider this may not reflect the future actual behavior with respect to adoption from shelters [7].

F. COVID-19 Pandemic and Public Interest in Pet Adoption

This study aims to define if the global interest in pet adoption increases after the pandemic declaration and if the effect has been sustainable. Moreover, the data were collected between 2015 and 2020. Eventually, the study concluded that in the early phases of the pandemic, the global interest in pet adoptions has surged. However, it was not sustainable. Following the COVID-19 pandemic, pets may face separation anxiety when their owners return back to work [8].

G. Exploring User Information Needs in Online Pet Adoption

Profiles

This study demonstrated how important to understand adopters’ needs to provide information about pets through analyzing user needs to determine the kinds of information required when searching for a new pet, specifically a dog or cat. Furthermore, Study participants rated several physical and behavioral characteristics based on their significance level. In general, the study shows cat adopters have an interest in cats’ personalities and behavior. On the other hand, dog adopters are interested in dogs’ physical characteristics [9].

H. Shelter Operations: Pet-Friendly Shelters

The study focused on the idea that pet-friendly shelters are most frequently organized by either local animal control offices or county/state animal response teams. The main idea is about sheltering operations involve endangered people who own pets, but most emergency shelters don’t accept pets due to health and safety regulations. If there is no opportunity to bring their pets with them to safety, some pet owners will refuse to vacate or will delay vacating. Pet-friendly sheltering is one of the most concerted methods of providing emergency accommodations for pet owners and their pets. Furthermore, it is a public human emergency shelter that is located within the same area. Eventually, the presence of pet-friendly shelters can increase the likelihood that endangered pet owners will evacuate to safety with their animals during an emergency [10].

I. As animal shelters fill up, new technology helps reunite

lost pets with owners

Sammie wrote about how technology helped pet owners find their animals in shelters with the help of the Petco Love


3 | P a g e


website service; Petco Love lost that use of face recognition technology. This technology helps pet owners search for their lost pets by uploading a picture of the pet on the website and looking for a match or the shelter, or anyone who found a pet can upload an image. However, not all animals are lost; some of them were abended by their owners, in addition, no one will look for them [11].

J. Pets and the Net: Helping Animals in Need

This article discusses how the internet can help adopt a pet with a particular condition, a specific color uncommon for someone to adopt, or any pet. By using the internet, shelters and pet owners can help animals in need and build awareness through social media. Some shelters use an emotional method like posting as if they were the animal, and this animal is desperate looking for a home and loving family. However, some shelters might have a low profile; thus, this article’s author provided suggestions like using well-timed hashtags, pet-based influence marketing, Etc. [12].

IV. SIMILAR SYSTEMS

Our benchmark is the “Petfinder” website [13]. The main reference for our project, as we aspire to be the best by providing similar services to theirs, but in a better and more distinctive way so that we are distinguished by services that are not available anywhere and specific to our platform. Furthermore, to be the trusted reference for the Saudi community in promoting the culture of animal adoption. Furthermore, many websites have a similar purpose to our platform and are available online will be mentioned as follow:

A. Adopt a Pet

Adopt a pet is a nonprofit website that helps adopt animals from different shelters and rescues. The website has a simple layout for novice users, and you can find the type or kind of animals you want and filter more personal performance like breed and age [14].

B. The Shelter Pet Project

The Shelter Pet Project is a cooperative effort between the two highest animal welfare organizations, Maddie’s Fund and Humane Society of the United States. Their purpose is to make shelters the first place for adopters to get a new pet, guaranteeing that all pets find loving and caring homes [15].

C. Petango

Partnered with animal welfare organizations across Canada and U.S., Petango is the first adoptable pet search site to offer real-time updates of adoptable pets in shelters exclusively [16].

Eventually, All the mentioned websites will be considered in our project implementation. Furthermore, we will do our best to overcome all the possible functional, and nonfunctional issues found and provide high-quality services for our end-users. Table 1 below demonstrates the uniqueness of our platform against the mentioned websites.

TABLE I. PROPOSED PLATFORM UNIQUENESS

Feature /

Website

Platform Uniqueness

Petfinder Adopt a

Pet

The

Shelter

Pet

Project

Petango Leen

Looking for a

pet by location

Ease of adoption

process

Able instant messaging

between users

Provide

discussion forums

Provide

donation

associations and shelters

Looking for veterinary

Clinics by

location

Provide all

services free with no fees

Provide pet

delivery service

V. MOTIVATION

Shelters are not found in abundance; they are expensive and not well funded. Mostly they start with a group of volunteers, and if they had good funding from a charity or an organization, they might be able to run a shelter. The money is used to maintain the building, cover health bills, food, and equipment. However, those shelters rarely have good funding because they can’t set up a clear plan for the expenses. Currently, there is a large movement and a better understanding of animal adoption than it was before. Therefore, our website is targeted for that reason. Providing a platform for people to be engaged with animal adoption in the comfort of their homes will cause in decreasing the load on the shelters. By adopting an animal, you’re making room for another to be taken in. In worse cases, you will be rescuing abandoned animals off the streets or animals that are in need of care. “Leen” aims to increase the awareness of the “Adopt don’t shop” campaign, instead of paying breeders to buy a pet, you can pay less to rescue an animal that needs a loving home and give more support to spread awareness of adoption.


4 | P a g e


VI. METHODOLOGY

“Leen” platform will adopt and follow the waterfall approach model, which go through logical and sequential stages, so those clear objectives are defined for each stage of software development and must be completed in each stage. Then, move to the next stage without considering the previous stages after completing them. The waterfall model consists of several phases (requirement analysis, planning, design, implementation, and testing). Our system requirements are clear, precise, and static. Therefore, the waterfall approach is the most suitable to use. The platform will apply the following tools, programming languages, and techniques during the implementation phase:

• Front-end: HTML, CSS, JS, Bootstrap3, Sweetalert2.

• Back-end: Python, Flask, Twilio, phpMyAdmin, XAMPP.

• UI Mockups: Axure.

VII. ARCHITECTURE

Fig. 1 represents the data flow diagram level 1 that includes the core processes exists in the system, and there are ten processes complete, namely:

• 1.0 Login/signup, it takes the user data and verifies them to get access rights. The user data is checked and updated in the database.

• 2.0 Place rehome ad, after logging in the adoptee user can place a rehome ad for their pet.

• 3.0 Browsing categories, the user can browse in the ads categories.

• 4.0 Display ad, after a user selects an ad, it is displayed to be viewed and requested.

• 5.0 Submit a request, the submission is stored in the database waiting to be accepted or rejected.

• 6.0 Requests responses, the user can check their requests’ status and select them to be viewed.

• 7.0 View requests, view the selected request.

• 8.0 View clinics information, the admin can access and view the clinics information.

• 9.0 Manage users, the admin can access users and manage their data.

• 10.0 Notify users, the admin can manage and send the notifications that users receive.

VIII. IMPLEMENTATION

The following figures illustrate the platform’s interfaces, which are divided into three. Common interfaces accessed by everyone including the platform’s visitors, user interfaces accessed by platform’s users who have accounts, and admin interfaces accessed by the admin only.

A. Common interfaces

• Homepage.

In Fig. 2, this interface is called in the platform “Home”, and it will introduce the user to view “Leen” platform in general. A menu containing a set of options will be located at the top. On the menu, when a user clicks on option, the selected option will be highlighted, and the font will become bold. Further, when a user clicks on “Adopt your pet today !” button, “Look for a pet” interface will be appeared. A search bar will be shown in all interfaces to facilitate the searching process for users.

Figure 1: Platform Architecture


5 | P a g e


• About Leen, Services, Contact Us and Login.

In Fig. 3, this interface is called in the platform “About Leen”,

it will explain to the user the reasons behind the exitance for

“Leen” platform. Moreover, for Fig. 4, this interface is called

“Services”, and it will display the main services provided by the

platform. Further, Fig. 5 illustrate “Contact Us”, and it will

provide all the needed information to let the user contact and

reach us easily.

B. Admin interfaces

• Admin Control panel

For the below interface Fig. 7, the admin can view the platform

statistics and the status per region by providing quantitative

results, get access to all the platform’s services and functions,

view and analyze the performance, and track users’ activity

live, Etc.

Figure 2: Homepage interface

Figure 3: About Leen interface

Figure 2: Services interface

Figure 5: Contact Us interface

Figure 6: Login interface

Figure 3: Control panel interface


6 | P a g e


C. User interfaces

• Look for a pet service, and adoption request application.

In Fig. 8, this interface provides “Look for a pet” service, the user must answer all the questions and all adoption offers match user’s preferences will be shown. In Fig. 9, this interface will be displayed to the user who wants to adopt a pet from a specific offer and it contains a form that must be filled.

IX. CONCLUSION AND FUTURE WORK

To conclude, this paper presents the work performed in developing a web-based platform that supports the idea of using technology for a pet adoption process in Eastern province in Saudi Arabia specifically. In addition, the platform was designed to provide easy and quick services for pet adoption and to have it accessible to all interested people, services like discussion forums, instant chatting, pet delivery, and nearby veterinary clinics. Furthermore, there are several

recommendations, which might provide some enhancements to our platform in the future, for example expanding to be used in all Saudi Arabia regions, developing an application, and adding a section focusing on providing educational courses related to pets for whom interested in this field.

REFERENCES

[1] “Free pet adoptions study results,” Maddie's Fund, Nov-2012. [Online]. Available: https://www.maddiesfund.org/free-pet-adoptions-study-results.html.

[2] “ ‘Adopt, don’t shop’ number one motto for Saudi pet shelters”, Arab News, 2022. [Online]. Available: https://www.arabnews.com/node/2032371/saudi-arabia.

[3] A. Southland, S. Dowling-Guyer, and E. McCobb, “Effect of visitor perspective on adoption decisions at one animal shelter,” Journal of Applied Animal Welfare Science, vol. 22, no. 1, pp. 1–12, Mar. 2018.

[4] A. Bogle, “How social media helps bring shelter animals out of the shadows,” Mashable, 10-Aug-2016. [Online]. Available: https://mashable.com/article/social-media-shelter-animals.

[5] P. Jeyaraj and A. Aponso, “A review of techniques for image classification to enhance online animal adoption speed,” Proceedings of the 2020 12th International Conference on Computer and Automation Engineering, Feb. 2020.

[6] M. L. Taborda, M. Lemos, and J. J. Orejuela, “The Impact of Adopting a Pet in the Perception of Physical and Emotional Wellbeing,” ResearchGate, vol. 10, no. 2, pp. 53–74, Jun. 2019.

[7] J. M. Frank and P. C. Frank, “Attitudes and Perceptions Regarding Pet Adoption.” faunalytics.org, 2008.

[8] J. Ho, S. Hussain, and O. Sparagano, “Did the COVID-19 pandemic spark a public interest in pet adoption?,” Frontiers in Veterinary Science, vol. 8, May 2021.

[9] Z. M. Becerra, S. Parmar, K. May, and R. E. Stuck, “Exploring user information needs in online pet adoption profiles,” Proceedings of the Human Factors and Ergonomics Society Annual Meeting, vol. 64, no. 1, pp. 1308–1312.

[10] “Shelter Operations: Pet-Friendly Shelters.” SAMHSA, US, 2021.

[11] S. Purcell, “As animal shelters fill up, new technology helps reunite lost pets with owners,” reporter newspapers & Atlanta Intown, 08-Jul-2021.

[12] D. Hughes, “Pets and the net: Helping animals in need: Blog: Online Digital Marketing courses,” Digital Marketing Institute, 18-Feb-2020. [Online]. Available: https://digitalmarketinginstitute.com/blog/pets-and-the-net-helping-animals-in-need

[13] “Urgent Need for Pet Adoption - Find Dogs & Cats & More | Petfinder”, Petfinder. [Online]. Available: https://www.petfinder.com/.

[14] “Adopt a dog or cat today! Search for local pets in need of a home - AdoptaPet.com”, Adoptapet.com. [Online]. Available: https://www.adoptapet.com/.

[15] “The Shelter Pet Project”, The Shelter Pet Project. [Online]. Available: https://theshelterpetproject.org/.

[16] “Petango.com Online Pet Adoption & More. Welcome a homeless pet into your home today.”, Petango.com. [Online]. Available: https://www.petango.com/.

Figure 8: Look for a pet interface

Figure 9: Adoption request application interface

https://www.maddiesfund.org/free-pet-adoptions-study-results.html

https://www.maddiesfund.org/free-pet-adoptions-study-results.html

https://www.arabnews.com/node/2032371/saudi-arabia

https://www.arabnews.com/node/2032371/saudi-arabia

https://mashable.com/article/social-media-shelter-animals

https://digitalmarketinginstitute.com/blog/pets-and-the-net-helping-animals-in-need

https://digitalmarketinginstitute.com/blog/pets-and-the-net-helping-animals-in-need

https://www.petfinder.com/

https://www.adoptapet.com/

https://theshelterpetproject.org/

https://www.petango.com/


1 | P a g e

Published In: IEEE 2nd Asian Conference on Innovation Technology

Road Damages Detection and Classification

using Deep Learning and UAVs Mohammad Aftab Alam Khan1, Mohammad Alsawwaf2, Basheer Arab3, Mohammed AlHashim4, Faisal Almashharawi5,

Omran Hakami6, Sunday O. Olatunji7, Mehwash Farooqui8 and Atta-ur-Rahman9

Department of Computer Engineering, College of Computer Science and Information Technology,

Imam Abdulrahman Bin Faisal University, P.O. Box 1982, Dammam 31441, Saudi Arabia

Email:[mkhan, mkalsawwaf, 2180007157, 2180002260, 2180007096, 2180005916, osunday, mfarooqui,

aaurrahman]@iau.edu.sa

Abstract— The Road health management is particularly important, especially for big cities and countries. Problems that occur on roads like road cracks can be extremely dangerous to drivers' and passengers' lives. In this paper, a road monitoring system is proposed to detect and classify the occurring problems on the road that happened due to obstacles, such as excavations. This work will help in repairing critical road damages faster and save people from accidents that are caused by these damages. The proposed model will detect and classify the problem related to road damage into categories (cracks, potholes, and other damages). This proposed model will be built using a deep learning technique which is convolutional neural networks (CNN). It has been found that CNN is widely used in this area and images detection and classification because it shows high performance. Numerous works have been done in this field, but it is hoped that this proposed technique will achieve better results. The proposed model will be connected with a drone, and it is linked to a web application to demonstrate the results and manage the system. Also, an announcement to government agencies such as the Ministry of Transportation or police could be sent using the web application. Theoretically, the outcomes from of this work shall demonstrate extremely reasonable findings in detecting and classifying road damages from real-time recording.

Keywords: Deep Learning, CNN, Classification, Detection, Road Damages

I. INTRODUCTION

Big cities depend on third parties and companies to maintain road safety. Although, there are few works and applications that identifies road problems quickly and efficiently. With the development and marketing of self-driving cars, road maintenance will become increasingly more crucial [1]. Most governments and road health monitoring authorities detect and classify road damage manually. Nevertheless, manual detection has apparent drawbacks, such as relatively low efficiency, high labour costs, and extremely sluggish computation and processing of numerous data [2]. Saudi Arabia is one of the authorities that depend on a manual approach for detecting and classifying damages on roads. Developing a system that handles this kind of problem using AI technologies will help in increasing the speed of response and efficiency in classifying the damages. Such work has not been conducted yet in the Kingdom of Saudi Arabia. This kind of work will be in line with and will improve the technological development adopted by the Kingdom's vision 2030.

A road collision is one of the leading causes of mortality worldwide. One of the contributing causes of a traffic collision is road damage [3]. With the existence of road damage, problems will exist and affect drivers, cars, and governments. To address this issue, early detection of road deterioration is necessary. For that reason, developing a method to examine roads health regularly would lead to fixing damages faster to overcome accidents because of these damages. This work aims to provide a reliable system used to provide routine checking on highways. Next, provide reports about the damage detected with location, image, and damage type for each damage.

A variety of vision-based damage detection approaches, mostly based on image processing techniques (IPTs) [4]. Yet, the use of such contextually (i.e., employing previous information) image processing is constrained since picture data captured in real-world circumstances varies greatly. Other works like the work in [4], [3], [5], and [6] used deep learning as an optimal solution because it is considered more adaptive to a real-world situation. However, some studies such as [5] propose that an approach of hyper red methods from both techniques will make an enormous difference and outperforms other experiments which are only conducted by one approach. In using the deep learning approach there are several ways to build the model. The most noticed methods are using a pre-trained model like YOLO in [3], [7] and [8] or building an algorithm with optimized parameters from scratch like CNN in [4] and [9]. Based on the review of related works of literature CNN deep learning algorithm is most used in this type of problem.

In this paper, a road monitoring system will be designed to detect and classify the occurring problems on the road that happened due to obstacles, such as excavations. This work will help in repairing critical road damages faster and save people from accidents that are caused by these damages. The model will detect and classify the problem related to road damage into


2 | P a g e


categories (cracks, potholes, and other damages). This proposed model is built using a deep learning technique which is convolutional neural networks (CNN). It has been found that CNN is widely used in this area and images detection and classification because it shows high performance. Python programming language will be used to train the model and to build the web application. A python library called Flask will be utilized to develop the web app and the APIs. The proposed model will be installed on a drone, and it is linked to a web application to demonstrate the results and manage the system. Also, an announcement to government agencies such as the Ministry of Transportation or police could be sent. Also, the dataset will be collected from online sources in addition to gathering some data of local roads using a drone. The proposed model should have reasonable accuracy in detecting damages. Add to that, a drone could be programmed to have autonomous navigation that will help in taking images and collecting data. For a user-friendly experience, that model will be integrated with a web application to manage the process and see the results.

This paper is structured as follows; The next section (section II) discusses a review of related works and studies. Section III states the methodology of the proposed system. Section IV talks about the functionalities and limitations. Section V is about the expected results and outcomes and section VI is the conclusion.

II. REVIEW OF RELATED WORK

As this research focuses on detecting and classifying the damages on roads, numerous related studies have been reviewed. These related works different on the used methods and results. It has been noticed that most of the reviewed works used CNN deep learning algorithm. While there are few others that that implied pre-trained models. For example, in the work in [3], the researchers present an autonomous pavement distress detection system based on the YOLO v2 deep learning framework. The dataset is formed of 9053 images that were captured with mobile cameras. These images are divided into 7240 images for training and 1813 images for testing. Images were obtained from seven different municipal administrations in Japan. For distress detection, YOLO v2 achieved an F1 score of 0.8780. For future improvements, the authors are considering using Google street-view images. Where article [10] suggested a sensor-based road health monitoring system. To identify the type of road, the system employs deep learning-based classifiers, which run on resource-constrained devices such as smartphones. In this work, the sensory data of diverse types of roads was performed using two vehicles. The researchers have taken certain convolutional layers in Deep Neural Network (DNN) to extract the spatial features. The algorithm and its variants have a training accuracy of 98%.

However, study [9] proposes a deep-learning-based fracture detecting technique. The method used to build the proposed method was ConvNets or convolution neural net (CNN). A quantitative assessment was performed on a dataset of square image patches composed of 500 images acquired using a moderate smartphone. The collected dataset is split into 64% as training samples, 16% for validation and 20% as testing samples. CNN achieved 89.6% in F1-score which is better than the other algorithms. The work in [11] introduces a real-time automated surveying system for collecting, classifying, and mapping image-based distress data. A qualitative methodology is considered for detecting fractures from gathered data using a convolutional neural network (CNN). The data collection was restricted to 1500 images of cracked asphalt pavement surfaces. A total of 1350 images for training and 150 images for testing. The proposed CNN was able to classify the cracks with an accuracy of 97% during the training. 

In [12], the authors propose road surface damage detection using fully convolutional neural networks (CNN) with semi-supervised learning. The dataset is collected with the help of a camera installed in the vehicle while driving. it consisted of 40,536 images. Data augmentation is applied to the training set, 20% of the training set before data augmentation is randomly taken for the validation set. The authors achieved 0.94 accuracies in total with the semi-supervised approach. Also, paper [13] proposing a deep convolutional neural network (CNN) called CrdNet for damage detection. The dataset consists of 7282 grayscale images, and it is collected using a special inspection vehicle. A 6550 of the data is used for the training and 732 for testing. The proposed work had a mean average precision of 90.92%. The work in [14] used several neural network types to detect and Classify Road damage. It takes the densely connected convolution networks to work as the backbone for Mask R-CNN to extract the image feature and the feature pyramid network to combine the multiple scales features. To generate the road damage region a region proposal network. To classify the road damage a convolutional neural network is used. The dataset size of 9053 was captured using a smartphone mounted on a car.

Additionally, [15] proposed that the UAV can be a useful tool for the collection of reliable information about road pavement. In this paper, videos were collected from UAV platforms to process the images to detect the pavement of the road. SVM method is used to evaluate good results by training the collected data and testing random data selected from the trained data. The accuracy reached using the test data up to 92%. In paper [16], a special UAV was used to allow real-time controlling in a specified area to detect the flooring begging that required maintenance. DWT and PCA algorithms were used for the post-processing of images procedures, and SVM (Support Vector Machine) algorithm is used to segment and classify the images. The best achieving classification accuracy was 99.38%. In [5], deep learning and multitemporal methods were developed for automatic detection using Unmanned Aerial Vehicle (UAV). The convolutional neural network (CNN) uses a segmentation method that measures any change that happened in that area. The dataset holds 91,595 images from scene videos. The quantitative result achieves a higher F-measure of 98.70%.  In another region, a deep learning approach is used for multiclass instance segmentation to detect concrete damage in [17]. Mask Region-Based Convolutional Neural Network is the algorithm used in this paper to manage the damage of the concrete and employ the image to detect and segment the defect. The result of the algorithm showed the classification of the damage was efficient, and the accuracy was 96.87% used on the picked images. The picked images were


3 | P a g e


obtained using a digital camera. The dataset contains 576 images for training and validation, and 144 for testing and 96 images were added to the testing dataset.

The study in [18], the researcher proposed crack detection using google street view. There are two datasets in this paper. The second dataset suffered from imbalanced data. The researcher solved the problem with a resampling technique. The VGG-16 model based on the CNN algorithm was used to build the crack detection. The VGG-16 model achieved a 98.9% accuracy with the first dataset. The first dataset was composed of 27,000 images and split into 17,000, 5000, 5000 training, testing, validation, respectively. Where in paper [19], the researcher proposed Convolutional Neural Network (CNN) model to detect pavement distress. The CNN model used two datasets. The second dataset has 91,280 images and it split into 85% and 15% for training and testing sets respectively. The model reached 83.8% accuracy over the testing set using the second dataset. 2000 images and 105 images were added later to the dataset. In [20], a multiple deep convolutional neural network model is proposed. The single-shot multibox detector (SSD) convolutional neural network model has reached the highest accuracy, reached 87.6%. The model uses 8000 crack images, divided into 4800, 1600, 1600 for training, validation, and testing respectively. 

In paper [1], the researchers proposed a deep learning model to detect the cracks in the road surface. The dataset contained a total of 14400 images with different properties. The number of images for all properties in training, testing, and validation sets was 9600,2400, and 2400 respectively. The model was trained with the original brightness and after changing the brightness and comparing the results. The FCN deep learning algorithm was used. The highest F1-score was 0.85 for the dataset with the original brightness. A deep learning model to detect damaged roads with smartphone images in [21]. The dataset is composed of 7231 in the training set and 1813 in the test set. Some of the classes in the dataset are imbalanced. So, data augmentation was used in the training set to solve the imbalanced data. The darknet53 model was used with the YOLO framework to build the final model. The model achieved up to a 0.62 f1-score. In article [22], the researchers proposed a deeper network to detect road damaged. In this paper, two experiments and networks were done. The CNN algorithm was used in both experiments. The first experiment with the second network obtained higher accuracy with 91.3%. The dataset used in the first experiment was composed of 100,000 for crack and 100,000 non-crack images in testing. The training set was randomly selected with 20,000 for crack and 20,000 for non-crack.

III. PROPOSED SYSTEM METHODOLOGY

Road damage detection and classification will take many resources such as time and effort in training the model. In our paper, we decided to use and implement one of the famous deep learning algorithms. The CNN algorithm will be used and see the performance of the algorithm as we decided in chapter 2.

A. Model Building and Evaluation

The CNN algorithm is chosen to train the model and build APIs to communicate with the drone. We will build the CNN model from the scratch. Also, we will try to use the weights in the pre-trained model ResNet as a top layer in the CNN model and analyse their performance to choose the one who obtain the best results in terms of confusion matrix, precession, recall, and accuracy. We need to further analyse in confusion matrix to have an idea about the number of TP, TN, FP, and FN. Furthermore, image processing techniques will be used before building the model. to make sure the images are free from noises such as paper noise, salt noise, and other types of noises. Because we found that in some literature review apply some of the techniques and the accuracy was better than before. Moreover, hyperparameter is an important process to obtain the best model with the optimal parameters. The CNN model will be trained on RTX 3600 ti GPU or Google Collaboratory.

B. Architectural Design Approach

The multilayer architecture design approach is important to implement the web application. This work architecture includes three layers: the presentation layer, application layer, and data layer.

• Presentation layer: this is the first and top layer in the web application. It provides a presentation service that is presented to the end-user through GUI.

• Application layer: this layer is the middle layer in the architecture. This layer provides the business logic of the application.

• Data layer: This layer comes after the application layer, which is concerned with the storage and retrieves the web application data in the database.

The advantage of using multilayer architecture is that it will improve scalability, security, and flexibility. Also, any damage that happens to any layer will not affect the other layers and the system. This will help the system to be more secure. Furthermore, the web application will be more stable, because when a new feature is added to the system will update it in one layer.

1) Architectural Design The architecture design for the entire web application system is shown in figure 1:

• Presentation layer: This layer provides controlling the drone and choosing to start and end points will be displayed in (the name of the page). Also, the reports generated from the drone and the history of the report are listed in the history


4 | P a g e


and pending pages. Furthermore, the user can see the road via the camera’s drone and see the detection and classification process.

• Application on layer: This layer will handle the API requests. The drone will communicate with the deep learning model through an API to detect and classify the road state. Also, if the drone detects damage in the road will generate a report and send it to the web application.

• Data layer: This layer will handle the storing and retrieval process to and from the database by using SSMS. It will help the web application to store the reports that come from the drone and store them in the database. then, retrieve the reports and display them in the web application. Also, it stores the old and new drones’ information.

Figure 1: Web Application Architecture Design

IV. APPLICATION FUNCTIONALITIES AND LIMITATIONS

The road monitor system is a useful system that will help the drivers and develop Saudi Arabia in terms of safety. It has many functions that will help to make the road more safety. The system is divided into three parts the deep learning model, drone, and website.

A. Deep Learning Model

The deep learning model is the main part of the paper. It will detect and classify the damage. The functionality of the model will be provided below:

• Detecting the damages in the road

• Classification of the damage type

Figure 2: Model Functionality

B. Drone

The live detection and classification will do using a drone. The drone is responsible for detecting and classifying road surface damage. In this part, we will list all the functionality for the drone below.

• Real-time detection and classification The drone is responsible for detecting and classifying the damaged road. This process is done by capturing the detected road. Then, the drone will communicate with the model throw an API to classify the image.

• Generating reports The drone will send the classified images to the website. Also, the drone will provide additional information with images such as location, and time.


5 | P a g e


`

C. Web Application

To see the classified images and make decisions to fix the road or not. The website will provide the capability to accept or reject the images provided by the drone. The functionality of the website will list them below.

• Define report status

After generating the report, the user will choose either to repair the road and it will send the request to the people who responded to fix it, reject if the damage not high, or misclassify the image.

• Controlling the drone

The website will provide a page to control the drone. The user will provide information like starting point and ending point, speed of the drone, and height of the drone.

• Monitor the drone

The user can monitor the road and the classification is done by the drone throw the website. Notification Once the drone generates the report the website will notify the user about the new reports from the drone.

Figure 5: Web-based System Diagram

Figure 3: Drone Diagram Figure 4: Drone Functionality


6 | P a g e


Figure 6: Web-based System Functionality

V. RESULTS AND DISCUSSION

In the road damage detection and classification paper, several findings and outcomes are generated because of this study. First, the work is a web application to build an interactive system with the user. Several user interfaces are built in the web application that made the system user-friendly, each interface is arranged to make the system clear for the user, and each interface will be appeared for the users according to his role in the system, the rest of the interfaces will be hidden and not accessible for the unauthorized user in his specific role. Several datasets are collected from different online resources. They will be reviewed to be used to train, validate, and test the model.

In the implementation process for the work, a reasonable accuracy is trying to be achieved for the model using the best algorithms with the proper dataset, and that to build a system that is able and classify the damages on the road, this is one of the aims of the research work. Also, a reliable system is needed to build the model on the drone and control it for the detection process. Also, the system generates reports for the detected damage that appeared on the user screen, and that will make the user able to fix the damage that is on the road.

VI. CONCLUSION

This paper proposes the detection and classification of road damages using deep learning. The system is comprised of a web application and a deep learning model that gives all the components of the system the ability to communicate. Training and testing the model will be done using the CNN deep learning algorithm. Furthermore, an online dataset is being prepared to be used for training the model. The web application, the drone, and the model will all be synchronized by using APIs such as Flask. These proposed techniques are hoped to achieve better results. Also, an announcement could be sent to government agencies such as the Ministry of Transportation and the Police. This research work will be based on real-time video recording, demonstrating extremely reasonable findings for detecting and classifying roads damages.

REFERENCES

[1] T. Lee, Y. Yoon, C. Chun, and S. Ryu, “Cnn-based road-surface crack detection model that responds to brightness changes,” Electron., vol. 10, no.

12, 2021, doi: 10.3390/electronics10121402.

[2] L. Wang, X. H. Ma, and Y. Ye, “Computer vision-based Road Crack Detection Using an Improved I-UNet Convolutional Networks,” Proc. 32nd

Chinese Control Decis. Conf. CCDC 2020, vol. 2, pp. 539–543, 2020, doi: 10.1109/CCDC49329.2020.9164476.

[3] V. Mandal, L. Uong, and Y. Adu-Gyamfi, “Automated Road Crack Detection Using Deep Convolutional Neural Networks,” Proc. - 2018 IEEE Int. Conf. Big Data, Big Data 2018, pp. 5212–5215, 2019, doi: 10.1109/BigData.2018.8622327.

[4] Y. J. Cha, W. Choi, and O. Büyüköztürk, “Deep Learning-Based Crack Damage Detection Using Convolutional Neural Networks,” Comput. Civ.

Infrastruct. Eng., vol. 32, no. 5, pp. 361–378, 2017, doi: 10.1111/mice.12263. [5] D. Han, S. B. Lee, M. Song, and J. S. Cho, “Change detection in unmanned aerial vehicle images for progress monitoring of road construction,”

Buildings, vol. 11, no. 4, pp. 1–14, 2021, doi: 10.3390/buildings11040150.

[6] S. Naddaf-Sh, M. M. Naddaf-Sh, A. R. Kashani, and H. Zargarzadeh, “An Efficient and Scalable Deep Learning Approach for Road Damage Detection,” Proc. - 2020 IEEE Int. Conf. Big Data, Big Data 2020, pp. 5602–5608, 2020, doi: 10.1109/BigData50022.2020.9377751.

[7] Y. K. Yik, N. E. Alias, Y. Yusof, and S. Isaak, “A real-time pothole detection based on deep learning approach,” J. Phys. Conf. Ser., vol. 1828, no.

1, pp. 1–8, 2021, doi: 10.1088/1742-6596/1828/1/012001. [8] H. Maeda, Y. Sekimoto, T. Seto, T. Kashiyama, and H. Omata, “Road Damage Detection and Classification Using Deep Neural Networks with

Smartphone Images,” Comput. Civ. Infrastruct. Eng., vol. 33, no. 12, pp. 1127–1141, 2018, doi: 10.1111/mice.12387.

[9] L. Zhang, F. Yang, Y. D. Zhang, and Y. J. Zhu, “ROAD CRACK DETECTION USING DEEP CONVOLUTIONAL NEURAL NETWORK Lei Zhang , Fan Yang , Yimin Daniel Zhang , and Ying Julie Zhu,” IEEE Int. Conf. Image Process., pp. 3708–3712, 2016.

[10] R. Mishra, H. P. Gupta, and T. Dutta, “A Road Health Monitoring System Using Sensors in Optimal Deep Neural Network,” IEEE Sens. J., vol. 21, no. 14, pp. 15527–15534, 2021, doi: 10.1109/JSEN.2020.3005998.

[11] M. M. Naddaf-Sh, S. Hosseini, J. Zhang, N. A. Brake, and H. Zargarzadeh, “Real-Time Road Crack Mapping Using an Optimized Convolutional

Neural Network,” Complexity, vol. 2019, 2019, doi: 10.1155/2019/2470735. [12] C. Chun and S. K. Ryu, “Road surface damage detection using fully convolutional neural networks and semi-supervised learning,” Sensors

(Switzerland), vol. 19, no. 24, pp. 1–15, 2019, doi: 10.3390/s19245501.


7 | P a g e


[13] J. Jabez and B. Muthukumar, “A CNN-Based Length Aware Cascade Road Damage Detection Approach,” Procedia - Procedia Comput. Sci., vol.

48, no. Iccc, pp. 338–346, 2015, [Online]. Available: http://dx.doi.org/10.1016/j.procs.2015.04.191. [14] Q. Chen, X. Gan, W. Huang, J. Feng, and H. Shim, “Road damage detection and classification using mask R-CNN with DenseNet backbone,”

Comput. Mater. Contin., vol. 65, no. 3, pp. 2201–2215, 2020, doi: 10.32604/cmc.2020.011191.

[15] F. Dadrasjavan, N. Zarrinpanjeh, A. Ameri, G. Engineering, and Q. Branch, “Automatic Crack Detection of Road Pavement Based on Aerial UAV Imagery,” Prepr., no. July, pp. 1–16, 2019, doi: 10.20944/preprints201907.0009.v1.

[16] V. Barrile, E. Bernardo, A. Fotia, G. Candela, and G. Bilotta, “Road safety: Road degradation survey through images by UAV,” WSEAS Trans.

Environ. Dev., vol. 16, pp. 649–659, 2020, doi: 10.37394/232015.2020.16.67. [17] P. Kumar, A. Sharma, and S. R. Kota, “Automatic Multiclass Instance Segmentation of Concrete Damage Using Deep Learning Model,” IEEE

Access, vol. 9, pp. 90330–90345, 2021, doi: 10.1109/ACCESS.2021.3090961.

[18] M. Maniat, C. V. Camp, and A. R. Kashani, “Deep learning-based visual crack detection using Google Street View images,” Neural Computing and Applications. 2021, doi: 10.1007/s00521-021-06098-0.

[19] C. Zhang, E. Nateghinia, L. F. Miranda-Moreno, and L. Sun, “Pavement distress detection using convolutional neural network (CNN): A case

study in Montreal, Canada,” Int. J. Transp. Sci. Technol., no. xxxx, 2021, doi: 10.1016/j.ijtst.2021.04.008. [20] X. Feng et al., “Pavement Crack Detection and Segmentation Method Based on Improved Deep Learning Fusion Model,” Math. Probl. Eng., vol.

2020, 2020, doi: 10.1155/2020/8515213.

[21] A. Alfarrarjeh, D. Trivedi, S. H. Kim, and C. Shahabi, “A Deep Learning Approach for Road Damage Detection from Smartphone Images,” Proc. -

2018 IEEE Int. Conf. Big Data, Big Data 2018, pp. 5201–5204, 2019, doi: 10.1109/BigData.2018.8621899.

[22] L. Pauly, H. Peel, S. Luo, D. Hogg, and R. Fuentes, “Deeper networks for pavement crack detection,” ISARC 2017 - Proc. 34th Int. Symp. Autom.

Robot. Constr., no. Isarc, pp. 479–485, 2017, doi: 10.22260/isarc2017/0066.


1 | P a g e

Published In: 12th International Conference on Computer Science and Information Technology

(CCSIT 2022)

A COMPARISON BETWEEN VGG16 AND XCEPTION

MODELS USED AS ENCODERS FOR IMAGE

CAPTIONING

Asrar Almogbil1,2, Amjad Alghamdi1, Arwa Alsahli1, Jawaher Alotaibi1

Razan Alajlan1, Fadiah Alghamdi1

1 Department of Computer Science, college of Computer Science and Information

Technology, Imam Abdulrahman Bin Faisal University, Dammam, Saudi Arabia

2 Department of Computer Science, college of Computer Science and Information

Technology, Imam Abdulrahman Bin Faisal University, Dammam, Saudi Arabia

[email protected]

ABSTRACT

Image captioning is an intriguing topic in Natural Language Processing (NLP) and Computer Vision (CV).

The present state of image captioning models allows it to be utilized for valuable tasks, but it demands a

lot of computational power and storage memory space. Despite this problem's importance, only a few

studies have looked into models’ comparison in order to prepare them for use on mobile devices.

Furthermore, most of these studies focus on the decoder part in an encoder-decoder architecture, usually

the encoder takes up the majority of the space. This study provides a brief overview of image captioning

advancements over the last five years and illustrate the prevalent techniques in image captioning and

summarize the results. This research study also discussed the commonly used models, the VGG16 and

Xception, while using the Long short-term memory (LSTM) for the text generation. Further, the study was

conducted on the Flickr8k dataset.

KEYWORDS

Image Captioning, Encoder-Decoder Framework, VGG16, Xception, LSTM.

1. INTRODUCTION

One of the most challenging and important topics in computer vision and natural language

processing is image captioning [1],[2]. Image captioning aims to generate a natural language

description based on the association between the objects in the given image. Image captioning

can be helpful in different applications such as human-computer interaction and providing help

for visually impaired persons [3]. Therefore, several studies have developed an image captioning

model [4,5]. Initially, the studies related to image captioning were focused mainly on generating

natural language descriptions for video [6], following the studies describing neural caption

generation architectures [7, 8], such as the encoder-decoder architectures proposed in [9].

Recently, the encoder-decode architecture has shown much improved outcomes in efficiently

generating natural language descriptions of an image [10]. At first, the CNN layers are used to

extract the features of the image. Then the collected features are used by the Recurrent neural

network (RNN) model to attain the information from the image [11].

This study reviews the current advancement of image captioning models and summarizes the

underlying framework. Although much attention has been paid to the decoder, there has not been

enough focus on the encoder. To fill this gap, this study will compare the performance of two

different encoder models, namely: VGG16 and Xception. Moreover, a comprising that focus


2 | P a g e


(CCSIT 2022)

mainly on the performance of two widely used encoder - VGG16 and Xception is poorly

investigated, which will help further researchers to decide on the encoder model.

The rest of the paper is organized as follows. Section 2 presents related work. Section 3 discusses

the materials and methods used in this work. Experiments done are described in Section 4. The

result obtained is illustrated in section 5. Conclusions and future work in Section 6.

2. RELATED WORK

In this section, we will summarize multiple related studies from different sources. The studies

will be organized in chronological order ascendingly. The purpose of the related work is to gain

an understanding of the published studies relevant to the image captioning field.

In [12], they used the MSCOCO dataset and LSTM to encode the text and used CNN as an image

encoder to extract features, and they obtained the best result compared with their benchmark.

Another study [13] used VGG16 as an encoder, which aids in creating image encodings. Then,

the encoded images are fed into an LSTM. The proposed model was enhanced with hyper-

modifying parameters. As a result, the model's accuracy increased to attain state-of-the-art results.

In [14], different models of image captioning were used. A merge architecture was applied in this

study. CNN-5, vgg16, and vgg19 are the different CNN that are used along with the LSTM. The

experiment is done on Flickr8K dataset. A Bilingual Evaluation Understudy (BLEU) evaluation

metric is used to evaluate the models. The result showed that VGG16 is perform better than other

models. The authors in [15] compared different models of image captioning. All models were

conducted on the Flickr8K dataset. The architecture used in this study is encoder-decoder

architecture. For the encoder, two different CNN models are used, which are VGG16 and

InceptionV3. For the decoder part, two types of LSTM were used. The first type is a unidirectional

LSTM that works in one direction. The second type is bidirectional LSTM which works in two

directions. The proposed models used greedy search and beam search algorithms to generate the

captions. The results show that the InceptionV3 with bidirectional LSTM with beam search gave

the best result. The evaluation metric used is BLEU. In [16], the study proposed an image caption

generator in the Bengali language using a merged dataset of two languages by combining flickr8k,

BanglaLkey, and Bornon datasets. The transform-based and visual attention approaches were

used to implement the proposed model. The Transform-based approach used an inceptionV3

encoder and fed to a dense layer that contains an activation function. The visual attention approach

implements an Encoder-decoder framework as well. In the encoder part, the InceptionV3 and

Xception models were used. For the evaluation of the proposed model, the BLEU and Metor were

used.

In [17], the study proposed an image captioning model to use the model on any website to generate

the description of the inputted image. The proposed model followed the CNN-LSTM concepts

and was conducted on the flicker8k dataset. In [18], the study used CNN and RNN models, and

the Xception was trained using the flickr8k dataset. Another study used the xception model

coupled with LSTM in [19] to discover the object found in the image, detect the relationship

among the objects, and generate the proper captions. This study was trained using the fliker8k

dataset. The criteria to evaluate the model was the loss value. In [20], the authors compared the

most popular CNN architecture: Xception, Resnet50, InseptionV3, Vgg16, and Densent201.

Along with the LSTM decoder. The comparison was done to see the effect of the performance by

implementing different encoder models. The study used flicker8K dataset. The evaluation of the

comparison was the loss value and the accuracy to compare the model's performance. The study

[21] proposed different CNN models VGG16, Xception, and inception coupled with bi-

directional layer RNN models for an enhanced image captioning model. The models were trained

using flicker30K and coco datasets. The BLUE score and training and loss are used to evaluate

each model.


3 | P a g e


(CCSIT 2022)

3. MATERIALS AND METHODS

This section includes the description of the dataset used in the study and the different encoders:

VGG16 and Xception. Finally, the decoder model.

3.1. Dataset pre-processing

The dataset used in this work is Flicker8k, and it is available on GitHub [22]. Flicker8k

consists of two folders, the first folder contains only images, and the second folder

contains a text file with the image descriptions. For the data pre-processing phase, we

start working on the text file and organize it by mapping the image ID to a list of five

corresponding descriptions. After that, we worked on data cleaning by making all letters

in lower case, removing all the punctuations, and removing words with one character (e.g.

‘A’). Lastly, we saved all changes made in a new text file.

3.2. The Encoder models

3.2.1. VGG16 model

VGG16 is one of the most preferred CNN models as it has a very uniform architecture. Simonyan

and Zisserman developed this model in 2014 [23]. It contains 16 convolutional layers. By having

this amount of layers, the complexity would increase compared to the initial versions of the CNN

architecture. In the below Figures, the size is proportionally getting reduced. The two layers are

convolutional, and the output of these two layers is 224x224, followed by the max-pooling layer,

and the final output after the max-pooling layer of size 2x2 and stride of 2 will be reduced to

112x112. Finally, we have three fully connected layers called dense. Figure 1 shows the

architecture of the VGG16 model.

Figure 1: VGG16 Architecture


4 | P a g e


(CCSIT 2022)

3.2.2. Xception model

The Xception model, also called “Extreme Inception” was proposed by Francois Chollet. It is a

kind of CNN model used to extract the features from the image. Also, it is an extension of the

inception model that is also considered a type of CNN model [24], but a better and enhanced

version by reversing some steps to be more efficient and easier to modify [25]. The Xception

model contains 37 layers [20]. The model uses the depthwise separable convolutional layers

approach, which divides the image into K input channel with depth equal to 1, then applies the

filter into each part with depth equal to 1, after that compressed all input channels space then

applying 1*1 convolutional. The accuracy of the Xception model considers the highest among

the CNN model in agreement with the LR in [15]. Therefore, it gives the best result compared to

the other CNN models. Figure 2 illustrate the layers of the Xception model.

Figure 2 : Layers of Xception Model

3.3. The Decoder model

For the decoder model, LSTM based model was used, which takes input from the feature

extraction model to predict a sequence of words, called the caption.

Because LSTM overcomes the RNN's constraints, LSTM is more effective and superior to the

regular RNN. With a forget gate, LSTM can keep relevant information throughout the processing

of inputs while discarding non-relevant information. It can process not only single data points but

also complete data sequences [26].

4. EXPERIMENTAL

For the experiments, our model follows the encoder-decoder framework. Therefore, we tested and

evaluated two different encoder models. Furthermore, we illustrated the conducted processes for


5 | P a g e


(CCSIT 2022)

developing the models for each model and how we trained the models. Whereas the decoder

remains fixed during the experiment, as mentioned before, in order to focus on comparing the

performance of the encoder model.


6 | P a g e


(CCSIT 2022)

4.1. The encoder

In the feature extraction step, the size of the image features is 224x224. Extracting the features of

the image is done before the last layer. The goal of the last layer is to predict the classification of

an image. For this reason, the last layer is dropped. The models were trained on Flickr8k dataset

as was described in Section 3.

4.1.1 VGG16

• Before optimization

When we started the model's training, we split the dataset into two parts. The first part is for

training, and the second part is for testing. Flicker8k dataset contains a file named

"Flickr_8k.trainImages.txt" that includes 6000 image ID; this file is used for the training part. The

training phase will be done in three steps. The first step, load the features extracted from the

VGG16 model. In the second step, we will initiate a dictionary that contains descriptions for each

image. The third step, create tokenizing vocabulary by using Keras, which provides the tokenizer

class, and it can do the mapping from the loaded description data. In this step, we need to fit the

tokenizer given the loaded photo description text. The create_tokenizer() function is responsible

for fitting the created tokenizer given the loaded photo description text. In addition, it's for

mapping each word of vocabulary with a unique index value.

• After optimization

To optimize the result and reduce the loss obtained, we implement Adam algorithm, which is an

optimizer that increase efficiency of neural network weights.

4.1.2 Xception

• Before optimization

Our CNN-RNN model consists of three main parts: feature extraction (encoder), sequence

processor, and decoder. In the experiment, we used images with a size equal to 299x299. In the

features extraction step, which is done before the last layer of the model, we got an 8091 feature

vector. In training, feature extraction is loaded to the model, and the dataset is divided into two

parts: training with 7091 images and testing with 1000 images. Then, we tokenized the vocabulary

by mapping each word with a unique index value, and each image will have a maximum length

of sentence equal to 31. After that, we created a data generator to train the model to yield the

image in batches.

• After optimization

The Adam algorithm was implemented to optimize the model to improve its performance.

5. RESULT AND DISCUSSION

In this study, a total of four models were tested and evaluated —VGG16, VGG16 with

optimization, Xception, and Xception with optimization. The criteria for the comparison are taken

to be the loss instead of the accuracy value, and the standard metric for comparison used here is

the BLEU score.


7 | P a g e


(CCSIT 2022)

Table 1: Evaluation Table

Model BLEU-1 BLEU-2 BLEU-3 BLEU-4

VGG16

Epoch= 100

Loss=3.0345

0.522997 0.279958 0.186401 0.079141

VGG16 with

optimization Epoch = 100

Loss= 3.3746

Optimizer= Adam

0.498937 0.251331 0.168155 0.068864

Xception

Epoch= 50

Loss= 4.3955

0.096406 0.031889

0.020180 0.004638

Xception with

optimization

Epoch= 50

Loss=3.3618

Optimizer= Adam

0.550791

0.309441 0.216791 0.105341

The above table shows each model's performance in terms of the BLEU score, testing loss of the

implemented models, and the number of epoch with the optimizer if used.

Our results demonstrate that Xception with optimization BLEU scores outperformed the other

three models. The highest BLEU score achieved in the study was 0.550791. Both Xception with

optimization and VGG16 before optimization have similar scores. However, the loss of VGG16

was less than Xception with optimization. The main motivation for using the adam algorithm was

to show a significant improvement in the runtime and memory consumption and increase the

efficiency of neural network weights, as mentioned in the previous section. The caption generated

from the Xception with optimization model gives the best probability and more accurate captions

(see Figure 6). In contrast, the captions generated by the other three models (Figure 3-5) were

long sentences compared to Xception with optimization. We can infer from the experiment that

when the sentences are long, the more probable to make mistakes. In most situations, we found

that the short sentences are sufficient to explain an image, whereas lengthier sentences frequently

contain duplicate information and grammatical errors. The main challenge was to reduce the loss

in Xception models, and after using the optimizer, the loss decreased. Yet, it remained higher than

the loss obtained in VGG16 before optimization (see figure 7). Hence, we observed that when the

number of the Epoch is increased, the number of loss models will increase in the Xception models

due to the small size of the dataset.


8 | P a g e


(CCSIT 2022)

Figure 3: VGG16 Before Optimization

Figure 4: VGG16 After Optimization

Figure 5: Xception Before Optimization


9 | P a g e


(CCSIT 2022)

Figure 6: Xception After Optimization

Figure 7 Testing Loss Curve for Xception Before and After Optimization.

6. CONCLUSION

In this study, we used an encoder-decoder framework that been used in the previous studies. We

evaluated two different encoder models for the purpose of comparing the VGG16 and Xception

encoder models. So far, no study has been published comparing these two models which will help

researchers figure out which model is outperforming the other. The outcome of the comparison

shows that the Xception model, when implemented adam algorithm, will generate the most

accurate caption compared to the other three models. Moreover, the study attempted to use

Flickr8k open-source datasets. Despite the precise caption achieved, there is still a need for a

larger dataset. A large dataset will enhance the model’s performance.

0

1

2

3

4

5

6

0 1 10 50 100

Loss

Number of Epochs

Xecption after optimazation Xecption before optimazation


10 | P a g e


(CCSIT 2022)

7. ACKNOWLEDGEMENTS

We would like to thank Ms. Asrar Almogbil for her cooperation on providing the instructions.

We also extend our appreciation to Dr. Nida Aslam and Ms. Abrar Alotaibi for their continuous

efforts in helping and answering our questions during the experiment.

8. REFERENCES

[1] Raimonda Staniut¯ e and Dmitrij ˙ Seˇ sok. A systematic literature review ˇ on image captioning.

Applied Sciences, 9(10):2024, 2019.

[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence

Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE

international conference on computer vision, pages 2425–2433, 2015.

[3] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose MF Moura, Devi

Parikh, and Dhruv Batra. Visual dialog. In ´ Proceedings of the IEEE Conference on Computer

Vision and Pattern Recognition, pages 326–335, 2017.

[4] A. Ramisa, F. Yan, F. Moreno-Noguer, and K. Mikolajczyk, “Breakingnews: Article annotation

by image and text processing,” IEEE Transactions on Pattern Analysis and Machine Intelligence,

vol. 40, no. 5, pp. 1072–1085, 2018.

[5] H. Ben, Y. Pan, Y. Li et al., “Unpaired image captioning with semantic-constrained self-

learning,” IEEE Transactions on Multimedia, vol. 24, pp. 904–916, 2021.

[6] Vinyals, Oriol, et al. ”Show and tell: A neural image caption generator.” Computer Vision and

Pattern Recognition (CVPR), 2015 IEEE Conference on. IEEE, 2015.

[7] M. Tanti, A. Gatt, and K. P. Camilleri, “What is the Role of Recurrent Neural Networks (RNNs)

in an Image Caption Generator?,” Aug. 2017.

[8] Sulabh Katiyar and Samir Kumar Borgohain. Comparative evaluation of cnn architectures for

image caption generation. arXiv preprint arXiv:2102.11506, 2021.

[9] Sutskever, Ilya, Oriol Vinyals, and Quoc V. Le. ”Sequence to sequence learning with neural

networks.” In Advances in neural information processing systems, pp. 3104-3112. 2014.

[10] F. Huang, X. Zhang, Z. Zhao, and Z. Li, “Bi-directional spatial-semantic attention networks for

image-text matching,” IEEE Transactions on Image Processing, vol. 28, no. 4, pp. 2008–2020,

2019.

[11] S. Li, Z. Tao, K. Li, and Y. Fu, “Visual to text: Survey of image and video captioning,” IEEE

Transactions on Emerging Topics in Computational Intelligence, vol. 3, no. 4, pp. 297–312,

2019.

[12] A. T. S. B. a. D. E. Oriol Vinyals, "Show and Tell: Lessons Learned from the 2015 MSCOCO

Image Captioning Challenge," IEEE TRANSACTION ON PATTERN ANALYSIS AND

MACHINE INTELLEGENCE, vol. 39, p. 12, 2017

[13] V. V. P. M. M. Ashish Pateria, "Enhanced Image Capturing using CNN," International Journal

of Engineering and Advanced Technology (IJEAT), vol. 8, no. 4, p. 6, 2019.

[14] A. a. D. S. a. Y. M. V. Jmail, "IMAGE CAPTIONING: TRANSFORMING SIGHT INTO

SCENE," International Research Journal of Modernization in Engineering Technology and


11 | P a g e


(CCSIT 2022)

Science, vol. 02, no. 06, pp. 54-66, 2020.

[15] S. Takkar, A. Jain, and P. Adlakha, “Comparative Study of Different Image Captioning

Models.” Fifth International Conference on Computing Methodologies and Communication,

India, 2021.

[16] F. M. Shah, M. Humaira, M. A. R. K. Jim, A. S. Ami and S. Paul, "BORNON: BENGALI

IMAGE CAPTIONING WITH TRANSFORMER-BASED DEEP LEARNING APPROACH,"

arXiv, p. 20, 2021.

[17] M. M. Patil, "Experiment based on Deep Learning: Image," INTERNATIONAL JOURNAL OF

CREATIVE RESEARCH THOUGHTS - IJCRT , vol. 9, no. 12, p. 6, 2021.

[18] N. L. C. K. Satyabrat Mandal, "Automatic Image Caption Generation System," International

Journal of Innovative Science and Research Technology, vol. 6, no. 6, p. 4, 2021.

[19] V. U. G. S. V. M. Megha J Panicker, "Image Caption Generator," 2021.

[20] S. R. Sahrial Alam, "Comparison of Different CNN Model used as Encoders for Image

Captioning," 2021.

[21] A. P. Yash Indulkar, "Comparative Study for Neural Image Caption Generation Using Different

Transfer Learning Along with Diverse Beam Search & Bi-Directional RNN," 2021.

[22] the dataset available in: https://github.com/goodwillyoga/Flickr8k_dataset

[23] A. D. Hussam, "Compressed residual-VGG16 CNN model for big data places image

recognition," in 2018 IEEE 8th Annual Computing and Communication Workshop and

Conference (CCWC), Las Vegas, 2018.

[24] M. j. Panicke, V. Upadhayay, G. Sethi and . V. Mathur, "Image Caption Generator,"

International Journal of Innovative Technology and Exploring Engineering (IJITEE), vol. 10, no.

3, p. 6, 2021.

[25] F. Chollet, "Xception: Deep learning with depthwise separable convolutions," in Proceedings of

the IEEE conference on computer vision and pattern recognition, 2017.

[26] P. G. Shambharkar, P. Kumari, P. Yadav, and R. Kumar, “Generating Caption for Image using

Beam Search and Analyzation with Unsupervised Image Captioning Algorithm.” Fifth

International Conference on Intelligent Computing and Control Systems, India, 2021.

https://github.com/goodwillyoga/Flickr8k_dataset


57 | P a g e Published In:

Smart Inventory System

AUTHOR’S NAME (Hussain Khoder, Hussain Alyahya, Ali Almuallim, Zeyad Alquaimi, Mazin Almohsin)

College of Computer Science & Information Technology/Imam Abdulrahman bin Faisal University, King Faisal Road, King

Faisal University City, Dammam 34212

Email: [email protected], [email protected], [email protected] , [email protected], [email protected]

Abstract: Traditional inventory system are now obsolete, the costs of resources and time that traditional

inventory systems carry outweighs the cheap money cost for operation and this is where the smart inventory

system can be beneficial, it might require a bit more money at implementation, but it will save way more time

than traditional inventory systems. The proposed project attempts to create a system that makes inventory

management easier by utilizing NFC cards and readers, as well as sensors that alert administrators to products

that have been removed from the shelves.

I. Introduction:

System inventories are a comprehensive resource to

access all information about items or resources that

an organization owns and have in storage, usually

you would want an information system that can

manage the system inventory but not many

organizations have this kind of luxury, most

organizations opt to do it in papers and the inventory

manager will try to track items’ location and available

items by paper and observation, an information

system that could manage an inventory would act as a

centralized resource that could list all items that are

being held in storage and give all kind of information

that is needed. In this paper we will discuss the

importance of having a system that utilizes

technology instead of the traditional way which only

utilizes paper.

II. Difficulties of Basic Inventory Systems

Operations:

Traditional or manual inventory systems were used

previously by companies until automated systems

came along and replaced them. Why were they

replaced? Traditional or manual inventory systems

take a long time to operate, and this is a waste of

time, every process in these systems such as

controlling, updating, and maintaining the system is

done manually by employees. This means that

manual inventory systems daily and frequently track

stock of the products and processes that take place

inside the inventory. In traditional or manual

inventory systems, the process of tracking products is

done by employees which means that the results

come from the employees counting, by recording all

the processes of removing products that occur in the

inventory manually and repeatedly. In this situation,

the inventory process is difficult for the employees

because every time a product is removed or a product

is added to the inventory, the employee will have to

manually record these operations and update it at the

end of working hours, which may cause data loss, and

cost employees or the company time and money.

(Jane, 2017)

III. Why we need automation in inventory

management?

With the presence of manual inventory systems in

companies, companies will face many problems,

losses, and challenges that will cause loss of time and

money. One of the problems that companies may face

with manual systems is an incorrect calculation of the

company's inventory, which may cause errors in the

rest of the operations. Inventory systems need

accuracy in inventory data such as the number of

existing products stock and products removed by

employees or customers. Automated systems work on

this function using technology and software to update

the system when the product is scanned on the sensor

and thus the manager or person responsible for

management will be Notified of this removal or

addition along with the subtracted or added auto-

generated new inventory stock. Unlike manual

systems, when the employee forgets to record the

removed product, the manager will expect that the

product is still in the inventory stock. Finally,

automated inventory systems update all items and

inventory data automatically at the end of working

hours, which is the opposite of manual systems that

require manual updating at the end of working hours.

Thus, the process of ordering new products will be







58 | P a g e Published In:

easier and faster for managers. (Jane, 2017) (Duff,

2022)

IV. Functions of a Smart Inventory That Could

Improve Inventory Management

We created a project that uses many technological

functions to help us create a smart inventory

management system which contains a variety of

technologies that contribute an increase in collected

data, thereby facilitating the management process;

gathering information is a fundamental method for

improving inventory management. As a result, the

smart inventory contains tools and devices that help

in storing information, such as NFC reader, which it

can read the NFC chips and retrieve information from

it, then send this information to the Database, where

all of the information about the items is stored, each

element and item in the stock have an NFC tag that

contain all the necessary information, all items in the

smart inventory counted even if an item was taken all

the data will be record with the information of the

person who took the item. The Database which

contains all recorded data is linked with website,

which helps in organizing all data collected and

managing information accurately and quickly.

Conclusion:

We can see an increase of efficiency after

implementing such projects, this is because

traditional inventories require a lot of time, time

could be wasted looking for the item in shelves or

room or even different warehouses that are in

different buildings or cities, a smart inventory system

would hold all the data which can help with finding

the item and saving the time. There will be an

increase in accuracy of the assets that are recorded,

and we can track all items, so they don’t get lost or

miscalculated. Different branches can also benefit

from such projects in many ways either direct or

indirect by that we can see that compared to the

benefits the cost is negligible.

References: Duff, J. (2022, Apr 1). Smart Inventory Management System – The

Key to Improving Business Efficiency. Retrieved from

mytechmag: https://www.mytechmag.com/smart-

inventory-management-system-the-key-to-improving-

business-efficiency/

Jane, M. (2017, Sep 26). Difficulties in Using a Manual Inventory

System. Retrieved from bizfluent:

https://bizfluent.com/info-7920237-business-rules-

inventory-system.html

Iqbal, R., Ahmad, A., & Gilani, A. (2014). NFC based inventory

control system for secure and efficient communication.

Computer Engineering and applications journal, 3(1),

23-33.

Cheng, R. S., Lin, C. P., Lin, K. W., & Hong, W. (2015). NFC

Based Equipment Management Inventory System. J.

Inf. Hiding Multim. Signal Process., 6(6), 1145-1155.

OTHER DISSEMINATIONS

(POSTERS, BOOK CHAPTERS, WORKSHOPS…)


1 | P a g e Published In: King Fahd University of Petroleum and Minerals at “women in science” workshop

The figure above describe the connection between each equipment

The figure above displays the Home interface that

will be shown to the user after registration/login

Conclusion

References :

E. A. Holzapfel, A. Pannunzio, I. Lorite, A. S. Silva de Oliveira and I.

Farkas, "Design and management of Irrigation Systems," ChileanJar,

2009. [Online]. Available:

https://scielo.conicyt.cl/pdf/chiljar/v69s1/AT03.pdf. [Accessed 1

October 2021].

O. Debauche, S. Mahmoudi, M. Elmoulat, S. A. Mahmoudi, P.

Manneback and F. Lebeau, "Edge AI-IoT pivot IRRIGATION, Plant

diseases, and pests identification," Procedia Computer Science, vol.

177, p. 48, 2020.

This figure displays the

Welcome interface that

will be shown to the user

when she/he open the

application

The project end product will be a device that is

connected to sensors to measure soil moisture and

temperature. In addition, the sensor data will be

saved in Firebase real-time database which will be

displayed on a mobile application, this product will

help reduce the water loss as well as taking plants

need for water into consideration.

Results

Intelligent Watering System (IWS) is developed to be

able to detect the soil moisture level and based on

the given percentage the system will automatically

open/close the water source. This system initially will

consist of a soil moisture and temperature sensors, a

water pump, and an Arduino kit. The soil moisture

sensor will measure soil moisture at the root zone, if

the soil is dry then the water source will switch on

automatically with the help of the Arduino kit, and

when the soil moisture sensor sense that the soil is

moist enough the water source will be switched off

immediately. Therefore, there will not be neither over

watering nor under watering of plants. Thus, there will

not be any waste of water especially if the weather

was rainy or humid. Additionally, an application will be

used to monitor the humidity, temperature, and the

soil moisture levels of the plants in all time. This will

help users to monitor their plants from far distances,

so they will know the status of their plants. All these

information comes from the sensors that will be located

inside the soil. Our system will help to reduce watering

problems and will let gardeners take better care of

their plants with less effort.

Introduction

Use of IoT to improve existing solutions

Reduce water usage

Improve the overall health of the plants

Increase the efficiency of gardening

Help gardeners take care of their plants

remotely

Objective

The IWS project was developed to improve the

existing watering system using IoT to enhance the

process of watering plants automatically. The

developing of this project was divided into three

phases:

1. Planning and Gathering Requirements: in this

phase the overall plan of the project is

constructed, and the requirements are identified

and gathered.

2. Design: in the second phase we clarified how the

overall system is going to be designed in order to

achieve the system requirements. This includes

the hardware devices and their connections and

the application used to monitor the hardware.

3. implementation: we divided this phase into two

sections: development of hardware and

development of software. The two sections were

developed simultaneously then they were

connected as the final end product. The hardware

aspect consists of a soil moisture and temperature

sensors, a water pump, and an Arduino kit.

Arduino IDE was used to receive the data from the

sensors and control the water pump accordingly.

In addition, it was used to send the data to

firebase. The application on the other hand was

developed using Android Studio, it receives the

data from Firebase then display it to the user.

Methodology

The Intelligent watering System will be able to

measure the temperature and humidity of the soil

and display it to the user through the mobile

application to provide more insight. He/she will

also be able to open the irrigation system either

automatically or manually

INTELLEGENT WATERING SYSTEM

Authors

Raweyah abdullah

Hadeel Al-otaibi

Ferdous Qabbani

Noof Alborai

Layan Alsahli

Supervisor: Dr. Mohammed AL Qahtani

CCSIT GRADUATION PROJECTS PROCEEDING 2022

Documents

Transcript of CCSIT GRADUATION PROJECTS PROCEEDING 2022