A Context-aware Travel Recommendation

System based on User Emotion and

Personality

UA Piumi Ishanka

SupervisorProf. Takashi Yukawa

A dissertation submitted in partial fulfillment ofthe requirements for the degree of Doctor of Engineering

Nagaoka University of Technology

June, 2018

Acknowledgment

I would like to thank everyone who helped me in writing the dissertation andwithout their continuous support and encouragement, writing and finishingthis dissertation would have been beyond my power.

First of all, I have to express my gratitude and sincere thanks to my su-pervisor professor Takashi Yukawa for his guidance, understanding, patienceand constant support. His invaluable help of useful comments, remarks, andengagement through the learning process throughout the research and thesishave contributed to the success of this research. As well, my sincere grati-tude goes to co-examiners, Prof. Koichi Yamada, Associate Prof. HirotakaTakahashi, Associate Prof. Tessai Hayama and Associate Prof. HirofumiNonaka for their valuable comments and suggestions which have led to im-prove the dissertation.

I would like to thank all the teachers who supported me throughout mycarrier in my journey at the Nagaoka University of Technology. Also, mysincere, heartfelt gratitude goes to Japanese Government for sponsoringme the MEXT scholarship during my studies at the Nagaoka Universityof Technology. Lastly, most importantly my most profound gratitudegoes to my family for their endless love and encouragement. To thosewho indirectly contributed to this research, your kindness means a lot to me.

Thank you very much.

ii

Abstract

The context-aware recommendation systems attempt to address the chal-lenge of recommending products or items that have the greater chance ofmeeting user interests with the highest relevance by adapting to users orproducts current contextual information. In the context-aware recommen-dation domain, many contextual features have been identified as contextualparameters such as companion in the movie recommendation domain, timeand mood parameters in the music recommendation, and weather, season,travel type & etc, in the travel recommendation domain. In this study, weapply context-aware recommendation in tourist destination recommenda-tion by introducing user emotion and user behavior as the contextual pa-rameters. The emotion is one of the popular contextual parameters whichis adapted by many recommendation system studies and triggered the ef-fectiveness of emotion in the recommendation process though few workshave arisen in tourist destination recommendation. To utilize user emotionand incorporate in recommendation process along with user behavior, inour study we proposed a travel destination recommendation system. Wecompare and clarify the effectiveness of using emotion and user behavior inrecommendation process using both filtering technique and context model-ing techniques. As the filtering technique, we used pre-filtering and as thecontextual modeling used tensor factorization. In the filtering approach, thesystem recommendation is implemented in contextual pre-filtering paradigmand uses the contextual information to select most relevant item x user datafor generating recommendations by using item-item collaborative filtering.Top five destinations were generated as recommendations for each individ-ual context as well combining both contextual parameters to compare theeffectiveness in the recommendation by using Loglikelihood similarity andSimple Weighted Average predictive rating calculation algorithm. In thecontextual modeling, we used CANDECOMP/PARAFAC(CP) Tensor Fac-torization model which uses ratings from M users for N items under Q typesof contexts as a three-dimensional tensor and generated the top five recom-mendations for each context. Thus, we introduced a new corpus with theemotion parameter by employing Semantic Analysis techniques for place

iv

recommendation due to the lack properly recorded dataset and used in therecommendation system implementation. In the process of deriving emotiontags, we used the text reviews collected from TripAdvisor and defined anemotion tag for each place based on the lexicon-based semantic classifica-tion. Both recommendation approaches with context outperformed with theselected contextual parameters and results of tensor factorization approachwith the two parameters proved higher effectivity in tourist destination rec-ommendation compared to other approaches (Mean Average Precision =81.59%). Moreover, we employed user personality in user profiles to analyzehow users emotions influence the decision-making process in the interactionof the recommendation system based on their personality by using userstwitter profiles to trace how exactly they react to their moods. Our studyfocuses on a challenging field, such as tourist destination recommendationwhile selecting the emotion and user behavior as contextual parameters andthe selected contextual parameters proved user satisfactions towards therecommendation generated by the system.

v

Contents

1 Introduction 11.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . 31.3 Organization of the Dissertation . . . . . . . . . . . . . . . . 3

2 Background and Related work 52.1 Context-aware Recommendation . . . . . . . . . . . . . . . . 52.2 Contextual Information . . . . . . . . . . . . . . . . . . . . . 62.3 Travel Recommendation Systems . . . . . . . . . . . . . . . . 72.4 Emotion in Recommendation Systems . . . . . . . . . . . . . 9

2.4.1 What is emotion? . . . . . . . . . . . . . . . . . . . . 92.4.2 The Unifying Framework . . . . . . . . . . . . . . . . 92.4.3 Emotion in Context-aware Recommendation Systems 11

2.5 Personality in Recommendation Systems . . . . . . . . . . . 112.5.1 What is Personality? . . . . . . . . . . . . . . . . . . . 112.5.2 Personality in Recommendation Systems . . . . . . . . 132.5.3 Personality Implication on Emotion . . . . . . . . . . 15

2.6 Affect and Personality Acquisition and Context-aware Datasets 162.6.1 Context-aware datasets with affect and personality . . 162.6.2 Acquisition Methods . . . . . . . . . . . . . . . . . . . 17

2.7 Recommendation System Techniques . . . . . . . . . . . . . . 182.7.1 Recommendation Techniques . . . . . . . . . . . . . . 18

2.8 User Profiling in Recommendation Systems . . . . . . . . . . 232.9 Evaluation of Recommendation Systems . . . . . . . . . . . . 232.10 Sentiment Analysis Techniques . . . . . . . . . . . . . . . . . 252.11 Closing Discussion . . . . . . . . . . . . . . . . . . . . . . . . 26

3 Deriving a Dataset for a Context-aware Travel Recommen-

dation 273.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.2 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . 28

vi

3.3 Background and Related Work . . . . . . . . . . . . . . . . . 283.4 Incorporating Emotion into Place Recommendation . . . . . . 31

3.4.1 Emotion tag acquisition . . . . . . . . . . . . . . . . . 313.4.2 Proposed system . . . . . . . . . . . . . . . . . . . . . 363.4.3 Predictive rating calculation (item-item collaboration) 38

3.5 Experimental setup and evaluation . . . . . . . . . . . . . . . 393.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423.8 Closing Discussion . . . . . . . . . . . . . . . . . . . . . . . . 43

4 An Analysis of Emotion and User Behavior for Context-

aware Recommendation Systems using Pre-filtering and

Tensor Factorization Techniques 444.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444.2 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . 464.3 Background and Related Work . . . . . . . . . . . . . . . . . 46

4.3.1 Context-aware recommendation . . . . . . . . . . . . . 464.3.2 Pre-filtering with context-aware recommendation . . . 464.3.3 Contextual modeling with context-aware recommen-

dation . . . . . . . . . . . . . . . . . . . . . . . . . . . 474.4 Proposed System . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.4.1 Context acquisition . . . . . . . . . . . . . . . . . . . . 484.4.2 Recommendation system implementation . . . . . . . 49

4.5 Experimental Set up and Evaluation Protocols . . . . . . . . 524.5.1 Experimental setup . . . . . . . . . . . . . . . . . . . . 52

4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584.8 Closing Discussion . . . . . . . . . . . . . . . . . . . . . . . . 58

5 Adapting User Personality and Emotion in Context-aware

Travel Recommendation 595.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595.2 Background and Related Work . . . . . . . . . . . . . . . . . 605.3 Proposed System . . . . . . . . . . . . . . . . . . . . . . . . . 61

5.3.1 Context Acquisition . . . . . . . . . . . . . . . . . . . 615.3.2 Recommendation System Implementation . . . . . . . 64

5.4 Experimental Set up and Evaluation Protocols . . . . . . . . 655.4.1 Experimental setup . . . . . . . . . . . . . . . . . . . . 655.4.2 Evaluation protocols . . . . . . . . . . . . . . . . . . . 67

5.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695.7 Closing Discussion . . . . . . . . . . . . . . . . . . . . . . . . 69

vii

6 Conclusion and Future Work 716.1 Discussion of Contributions . . . . . . . . . . . . . . . . . . . 716.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

viii

List of Figures

2.1 Contextual pre-filtering example . . . . . . . . . . . . . . . . 62.2 Plutchik’s wheel of emotions . . . . . . . . . . . . . . . . . . . 102.3 The unifying framework: The role of emotions in user inter-

action with a recommender system . . . . . . . . . . . . . . . 112.4 Contextual Modeling Approaches . . . . . . . . . . . . . . . . 23

3.1 Collaborative Filtering Process . . . . . . . . . . . . . . . . . 293.2 Pre-filtering Process . . . . . . . . . . . . . . . . . . . . . . . 303.3 Emotion Tag Creation . . . . . . . . . . . . . . . . . . . . . . 363.4 Emotion Gathering . . . . . . . . . . . . . . . . . . . . . . . . 363.5 Recommendation Process . . . . . . . . . . . . . . . . . . . . 373.6 User-item Matrix . . . . . . . . . . . . . . . . . . . . . . . . . 383.7 Sample User Behavior Actions . . . . . . . . . . . . . . . . . 393.8 Loglikelihood Similarity Values . . . . . . . . . . . . . . . . . 393.9 Recommendation List . . . . . . . . . . . . . . . . . . . . . . 393.10 Mean Average Precision Values with Emotion Groups . . . . 41

4.1 Collaborative Filtering Process with Pre-filtering Tech-nique.The filtered user-item matrix data are used to predictrating for a user . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.2 Tensor Factorization Model. . . . . . . . . . . . . . . . . . . . 484.3 Emotion Acquisition.Logged in users’ emotion were collected

in Plutchik’s emotion classification scale. . . . . . . . . . . . . 484.4 User Rating Capturing. User preference for a place gathered

in five-star scale. . . . . . . . . . . . . . . . . . . . . . . . . . 494.5 Emotion Tag Creation.The users were asked to tag an emo-

tion for place when they are logged into the system. . . . . . 494.6 Architecture for the Recommendation System with two Rec-

ommendation Approaches. The Tensor factorization followsthe CP factorization model while item-item collaborative fil-tering was used as the Pre-filtering technique. . . . . . . . . . 50

4.7 User-item Matrix Illustration . . . . . . . . . . . . . . . . . . 51

ix

4.8 Recommendation Lists. The recommended lists in the figureshow places provided by three recommendation approachesTFE,TFUB and TFEUB. . . . . . . . . . . . . . . . . . . . . 53

4.9 Average Precision Values with Emotion Groups . . . . . . . . 554.10 User Rating Behavior based on User Emotion . . . . . . . . . 57

5.1 User emotion classification from tweets. . . . . . . . . . . . . 615.2 IBM Bluemix to analyze personalities. . . . . . . . . . . . . . 635.3 User profile details for a user. . . . . . . . . . . . . . . . . . . 645.4 Proposed recommendation system. . . . . . . . . . . . . . . . 665.5 Recommended places for a user. . . . . . . . . . . . . . . . . . 665.6 RMSE values for CFUFE and CFN. . . . . . . . . . . . . . . 69

x

List of Tables

2.1 Travel Recommendation Systems with contextual parametersand recommendation details . . . . . . . . . . . . . . . . . . . 7

2.2 Personality Models . . . . . . . . . . . . . . . . . . . . . . . . 122.3 Personality based Recommendation . . . . . . . . . . . . . . . 142.4 Personality Implication on Emotion by Jonson (2009) . . . . 162.5 Personality and Emotion based Data Corpuses [95, 26] . . . . 17

3.1 Unigrams/Bigrams from Macquarie Thesaurus. . . . . . . . . 323.2 Wordnet Affect Lexicon. . . . . . . . . . . . . . . . . . . . . . 323.3 Terms from General Inquirer. . . . . . . . . . . . . . . . . . . 333.4 Sample results of tone analysis created by ToneAnalyzer . . . 333.5 Extract from the emotion lexicon used in the classification . . 343.6 Emotion tag derivation based on TF values . . . . . . . . . . 343.7 Emotion tag derivation of TFs . . . . . . . . . . . . . . . . . 353.8 Sample results of tone analysis created by ToneAnalyzer . . . 373.9 Precision and Recall matrix . . . . . . . . . . . . . . . . . . . 403.10 Precision and Mean Average Precision values . . . . . . . . . 413.11 User satisfaction . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.1 Example of the Results for Contextual Information for a User 524.2 Precision Values and Mean Average Precision Values . . . . . 554.3 Precision Values and Mean Average Precision Values . . . . . 574.4 User Satisfaction . . . . . . . . . . . . . . . . . . . . . . . . . 58

5.1 Examples from the NRC Emotion Lexicon. . . . . . . . . . . 625.2 Daily Emotion Derivation for a Tweet. . . . . . . . . . . . . . 635.3 Precision and Mean Average Precision . . . . . . . . . . . . . 685.4 User Satisfaction . . . . . . . . . . . . . . . . . . . . . . . . . 68

6.1 Personality implication on emotion . . . . . . . . . . . . . . . 73

xi

Chapter 1

Introduction

1.1 Motivation

A large number of information is available on the internet has created anecessity of filtering information for the various types of user needs whichcreate an obligation of delivering most accurate data, more efficiently by alle-viating the problem of information overload. The recommendation systemsreduce the information overloading and provide users with personalized con-tents and services by searching through a significant amount of dynamicallygenerating data. The researchers of the recommendation system domainhave proposed vital no. of techniques for last decades and it is not onlyproviding solution to the information overload but also creates various num-ber of opportunities and challenges for online businesses, government, andother domains to develop in theoretical aspect as well as in the practicalimplementation of recommendation systems with the advancement of var-ious technologies.Information Recommendation has become an extensivelygrowing research area since the 1990s after emerging the notion of recom-mendation systems. From then on most of the recommender system aredeveloped by modeling the user preferences in predictive models where thepreferences captured as ratings given to each item in the system.

Recently many researchers draw their focus on the systems that canprovide recommendations with the available contextual information. In-tegration of contextual data to the recommendation systems enhances thestability of the recommendation in the sense of adaptability of the users cur-rent situations. Thus, the context-aware recommendation systems (CARs)are introduced to provide users with more relevant information, unlike tradi-tional recommendation approaches. Most of the contextual information arecharacteristics of an activity such as location, time and some dynamic fea-tures of user profiles such as users emotional states. Accordingly, to variouspsychological studies, a place can make an impact on peoples memories, sen-

1

timents, and emotional well-being [23]. The Psychogeography introduced byDebored[30] where the academia tracked the influence of geographical envi-ronments on the emotions and behavior of an individual. Though the placeis commonly analyzed considering as an urban context. The impact on hu-man emotional well-being has already studied, no taxonomy has derived onemotion that a place may evoke in people[57].

The emotion in recommender as the context has been firstly introducedby Gonzalez et al. [84] by considering how emotion influence users decisionmaking at the interaction with the system. From then on, the emotion hasbeen a popular contextual parameter in various domains, especially in movieand music recommendations.

One of the key challenges in context-aware recommendation developmentis the unavailability of properly recorded data. The issue has created newresearch question in the context-aware recommendation of deriving new datacorpus and testing the feasibility in the recommendation. In travel recom-mendation domain non. of such context-aware datasets are exist. Thus, forthe proposed system, deriving emotional tags for the destination needed.

The user emotion collection as contextual data does as in any methoddescribed in affective computing state-of-the-art. These methods can be ei-ther explicit method or implicit method. The explicit methods are moreaccurate in extracting emotion where implicit methods are less burden theusers though less accurate. In adapting emotion in context-aware travelrecommendation, most importantly the appropriate emotion as a context isimportant. As well, the suitability of affect context with other contextualparameters is essential to investigate in providing more accurate service tothe users. Moreover, in emotion-based recommendation systems, thoughusers are recommended item based on their emotions how users may reactto make a decision may vary with their personalities. Psychologists haveproved that personality highly implicates emotion. Therefore, in improvingthe emotion-based recommendation, user personality also plays an impor-tant role. Most of the research work in the personality-based recommenda-tion, widely discuss the cold-start problem of collaborative filtering, grouprecommendation, and diversity in the recommendation. Apart from theseconcerns how user personality effect in emotion and rating behavior in therecommendation is also significant to investigate. In our study, the emotionas a context in travel recommendation empirically compared and further,the relation between user behavior and personality to recommendation taskin enhancing the accuracy of the proposed system is analyzed. The achievedresults showed that emotion in Tensor factorization techniques outperformswhile pre-filtering approach also performed better results compared to non-context approaches.

2

1.2 Research Objectives

Our study aims to trace the accuracy of emotion based context-aware recom-mendation by using pre-filtering and context modeling techniques in travelrecommendation. For such purpose, how user behavior and personality con-texts perform along with emotion in travel recommendation system wasanalyzed. For better recommendations, the following specific research goalswere defined.

Objective 01: Derive emotion included data corpus for travel recom-mendation. As mentioned early, lack of properly recorded data corpus totravel recommendation was one of the main concern of our study. In achiev-ing this, we used user reviews for travel destinations and derived emotiontags for each destination. Further, we testified the derived data corpus inpre-filtering techniques.

Objective 02: Analysis of user emotion and user behavior in travel recom-mendation by using pre-filtering and contextual modeling techniques.Whilecomparing emotion in two context-aware recommendation algorithm cate-gories; pre-filtering and tensor factorization we employed additional contex-tual parameter user behavior to compare and combine in proposed algo-rithms.

Objective 03: Adapt user personality and emotion in user profiles andanalyze the effect of emotion in a recommendation based on personality.

Adapting user personality is to extend our research problems in findingthe relation of emotion and user personality in travel recommendation sys-tem while users are interacting with the system. In achieving this goal, useremotion from users’ Twitter profile was abstracted and analyzed the text tohave emotion and personalities for the study.

1.3 Organization of the Dissertation

The work done in the thesis has resulted in several contributions to the stateof the art of Travel recommendation.

Chapter 02: A comprehensive related work on emotion-based recommen-dation and personality recommendation were discussed in chapter 2. Themotivation of the study by analyzing the importance of emotion as contextas well we discuss existing research work in Travel Recommendation, no-tions of the recommendation systems and context-aware recommendation,user profiling, semantic analysis techniques, etc. were further explained.

Chapter 03: In this chapter, the emotion tag derivation process by usinguser reviews for the selected destination was illustrated in detail includ-ing the implementation and obtained results of the proposed emotion-basedtravel recommendation using the derived datasets.

Chapter 04: An analysis of emotion as a context in collaborative filtering

3

as pre-filtering and Tensor factorization in contextual modeling discussed inchapter 4. The emotion, incorporated in each technique along with userbehaviour context as well as individually to testifying the recommendationaccuracy.

Chapter 05: In this chapter, further the user personality implication onemotion by user-user collaborative filtering was conferred.

Chapter 06: Finally, the concluding remarks of the study discussed in asthe last chapter by revisiting the research question and finally discuss futurework for the study.

4

Chapter 2

Background and Related

work

2.1 Context-aware Recommendation

Context-aware Recommendation Systems(CARs) suggest more relevant in-formation by considering the contextual information of the users since theusers preference in one context may be different from those in another con-text. For example, a user may buy a toy for his or her friends baby forbabys birthday but next time she will be willing to buy something for her orhim. By adapting context information to the recommendation system (RS)enhance the predictability of the suggestions effectively and able to providereliable output to the users.

The RS estimates the rating function R using a set of initialized ratingsthat either explicitly provided by the users or implicitly inferred by the sys-tem for user u and item i which is yet to be rated

R: User x Item �RatingIn a CARS, preferences are predicted by incorporating available contextualinformation into the recommendation process and R is estimated by

R: User x Item x Context �Ratingwhere user and item are the space of users and items and ratings is the spaceof ratings for the user and item pairs. The contextual information can beapplied in various stages of the recommendation process and the form of thecontext-aware recommendation can be contextual pre-filtering (contextual-ization of input), contextual post-filtering (contextualization of output), orcontextual modeling (contextualization of recommendation function) [4].

The contextual pre-filtering approach generate a recommendation in 2Dform of user x item by selecting the relevant data based on the contextwhich allows the recommendation algorithm to be any of the traditional

5

recommendation technique [3].

Figure 2.1: Contextual pre-filtering example

Moreover, filtering data according to the exact content can be too nar-row because certain aspect of overly specific may not be significant andmay not have enough data for the predictions in the recommendation pro-cess [2]. The Contextual post-filtering doesnt consider the context in theinput data and generate the recommendation same as traditional RS andgenerate the ranked list and use the context information to adjust the rec-ommendation list for each user either by filtering the recommendations fora given context or adjust the ranked list. E.g.: if suggesting a movie fora user in a weekend and he only watch comedies, the recommendation listcan be sorted out based on non-comedies films. The Contextual modelingapproach uses contextual information directly and predict users rating foran item. The contextual modeling introduces multidimensional recommen-dation functions which represents predictive models based on decision tree,regression, probabilistic models or heuristic calculations that use contextualinformation along with user and item data while some of them extend the2D recommendation function to multidimensional recommendation [4].

2.2 Contextual Information

The context has been defined in multiple ways and described as a multi-faceted concept in different research domains [82]. Most of the contextualinformation concerns characteristics of an activity such as location, timeand some dynamic features of user profiles, for example, users’ emotionalstate [31]. Hence, information e.g. location, time, social companion, andmood, etc., can be considered as the context in the case of recommendationsystems [22]. Most of the domains introduce context-awareness to increasethe efficiency and usability of information systems particularly when sucha system is accessed by mobile devices [62]. However, this is applicable toother domains as well, since usability of an information system is highly con-cerned in development process. The context-aware recommendation systemshave been categorized into two kinds: firstly, as systems that use contextualinformation as a criterion to filter items; and secondly, systems employingcontextual information at the time users can evaluate items devised by some

6

authors [65].

2.3 Travel Recommendation Systems

Petrevska et al. proposed a tourism recommendation system based on userpreferences, interests and desires, and suggested tourist attractions basedon profiling user behaviors such as reading other reviews to make decisions[76]. Destinations were observed as objects and changes in user preferencesfollowing subsequent visits tracked according to the rating and user behav-ior was considered in the profiling and recommendations generated based oncollaborative filtering. Sarkaleh M. et al. suggested a tourism recommen-dation system using location and user features as contextual parameters[84]. De Pessemier et al. suggested a system with group recommendationfor traveler destinations based on a users rating profile, personal interests,and specification for their next destination by following a hybrid approachin combination with content-based, collaborative filtering and knowledge-based strategies [29].

The following table summarizes the research works in context selectionsand the context-aware recommendation in Travel Recommendation.

Table 2.1: Travel Recommendation Systems with contextualparameters and recommendation details

Author Context Contextual Pa-rameters

Recommendation Pro-cess

Castillo [24] User Behavior Previously vis-ited places inuser profiles

Case-based reasoningand the K-Nearest,Neighbor-hood algo-rithm

Huang andBian [54]

User Behavior Location,Open Hours,Close Date,Mini TimeStay, AgeRange, Occu-pation

Bayesian networktechniques and ana-lytic hierarchy process

Gavalas andKenteris [39]

Location, time,weather, userbehavior

Location, time,weather

Collaborative filteringwhile considering con-textual information inpervasive environment

7

Yong, Robinand Bamshad[106]

User Behavior Trip Type,Trip dura-tion,OriginCity, Desti-nation,City,Month

CF with differentialcontext relaxation

Barranco,NogueraCastro, andMartinez [11]

Location andTrajectory

User’s speedand traveldirection

Improve recommen-dations by using acontext-aware filteringin CF

Noguera, Bar-ranco, Seguraand MartiNez[72]

Location Location Use both pre and post-filtering approaches

Soha,Taysirand Adel [69]

Use Behav-ior and UserMood

Weather, Timeof the day,Users LocationUser Mood:Happy, angry,Excited, TiredUser’s speedand traveldirection

Genetic Algorithmand Matrix factoriza-tion

Soe Tsyr Yuanand Chun-YaYang [103]

User emotion User behav-ior searchinghistory, desti-nation,stores,feedback ofemotionalwords

Use color imagery asthe uniform represen-tation for customers’expectation to facili-tate the scoring andranking

All the studies in the table attempted to incorporate various types ofcontextual information in the recommendation process. For example, SoeTsyr Yuan and Chun-Ya Yang et al in their work tried using emotion in therecommendation process, but their method to extract emotion was based oncolor and imagination [103]. Travel recommendation systems serve touristswith relevant and personalized destination suggestions for helping them tomake better decisions. State-of-the-art of recommendation systems can beanalyzed as web-based systems or mobile-based systems, while the web-based are being the predominant type. As well depending on the service,systems’ recommendation can be clustered. Suggested here are the destina-tion and construction of a complete tourist package, recommended suitableattractions at one specific destination, a detailed multi-day trip schedule,

8

and social capabilities [15].

2.4 Emotion in Recommendation Systems

2.4.1 What is emotion?

Emotions are mental states usually caused by an event of importance to thesubject [73]. A number of studies have been discussed to introduce basicemotions as theories such as Ekmans six basic emotions: joy, sadness, anger,fear, disgust and surprise. Plutchik proposed eight basic emotions includingEkmans six basic emotions, including trust and surprise as well while hefurther argued that the eight emotions form four opposing pairs namely joy-sadness, anger-fear, trust-disgust and anticipation-surprise as showed in thewheel of emotion in Figure 2.2 [70].

The emotion can be described by modeling it using the Universal modelor Dimensional model. The universal model describes emotion categories ashappiness, anger, sadness, fear, disgust, and surprise while the dimensionalmodel uses valance, arousal and dominance. Recently, more efforts havebeen made to use emotion in the recommendation process since emotionsare crucial factor in users decision making. Thus, researches have been con-ducted on individual domains or combining domains such as an affectivecomputing and recommendation systems. The role of emotion in the recom-mendation systems’ consumption chain has been discussed in three stages,namely the entry stage, the consumption stage and exit stage. All threewield great influence on human decision-making. Marko et.al proposed theUnifying Framework in the means of enhancing the recommender systemsusing affective data by identifying the position of the activity in the recom-mendation process which was later used in many research works [96].

2.4.2 The Unifying Framework

In using a recommender, users could induce by various stimuli that mayinduce the emotive state and influence their decisions at a considerable rate.Thus, considering how user is interacted with the system and how users con-sume the service, emotions play different roles. Based on how emotion maytempt in user interaction it is divided into three main categories namely (i).the entry stage (ii). the consumption stage and (iii). the exist stage (seeFigure 2.3). The Entry stage is the first phase of the proposed frameworkwhich considers the affective state of the user when he or she starts to usea recommender system which is stated as entry mood. This entry mood canbe caused users activities before interacting the system but still influencethe users choices when the recommenders system suggests a limited numberamount of contents items to a user. The Consumption stage states the ac-

9

Figure 2.2: Plutchik’s wheel of emotions

countability of emotive state of the user while interacting with the system asinduced by the contents consumption. As an example, the emotive responseto watch an image and watching a movie which leads to a single value ofemotive state or vector emotions varied over time. Thus, in the Exit stageusers emotive response right after finishing the content is exit mood whichwill influence users next activity of user interaction to the system.

The affective state of the end users has been detected as explicitly orimplicitly : the explicit detection is more accurate though intrusive whilethe implicit methods are less accurate though less intrusive since user isntaware while interacting the system [42].

10

Figure 2.3: The unifying framework: The role of emotions in user interaction

with a recommender system

2.4.3 Emotion in Context-aware Recommendation Systems

Adapting emotion as a context parameter to the recommendation systemwas first done by Gonzalez et al. (2007) [19]. From then on, many ap-plications arose on this topic due to the success of various research work.Emotion-based music recommendation is one of the domain both recom-mendation systems and affective engineering studies are merged [45, 87, 99].Due to the diversity and richness of music content and in context-based mu-sic recommendation systems, it often needs multidisciplinary efforts such asemotion description, emotion detection, etc., to achieve success. The movierecommendation domain has also been enriched by applying emotion to therecommendation and a few studies incorporated emotion in various stagesof movie consumption [109, 47, 67]. Thus, for an example emotion in eachstage is monitored and incorporated into the LDOS-CoMoDa dataset forthe recommendation process with reference to emotional contexts: Mood,DominantEmo and EndEmo [47].

2.5 Personality in Recommendation Systems

2.5.1 What is Personality?

The psychology concept of personality accounts for the individual differ-ences in our enduring emotional, interpersonal, experiential, attitudinal and

11

motivational styles and supposed to be stable across longer period [25]. Re-search works have proven that personality related with the users real-lifepreferences and this creates a good possibility to model long-term user pref-erences [80, 94].

Thus, users with different personalities incline to choose different kindsof content and this relation is domain dependent. In designing a recommen-dation system these relations are very valuable. Various studies have useddifferent types of personality models in their studies and the following tableshows some of such models and these models provide a scientific approachin describing and predicting individual differences [58].

Table 2.2: Personality Models

Author Model Domain Traits

Tomas et.al[25]

Five FactorModel

General Openness, Conscien-tiousness, Extraver-sion,Agreeableness,and Neuroticism

John [48] Four Tempera-ments

General Choleric, Sanguinic,Melancholic andPhlegmatic

Bart [89] RIASEC Vocational Realistic, Investiga-tive, Artistic, Social,Enterprising andConventional

Richard et.al[37]

Bartle types Video games Killers, Achievers, Ex-plorers and Socializers

Kenneth [92] Felder andSilvermanLearning StyleModel

Learning styles active/reflective,sensing/intuitive,visual/ verbal, sequen-tial/global

Sandraet.al[66]

Thomas-Kilmannconflict model

Group/conflictmodeling

Assertiveness, Cooper-ativeness

Among all these personality model, Five Factor Model or/and Big Five(FFM) became the most famous model. These two terms have been used in-terchangeably though they have been developed independently. Both thesemodels use the same factor labels despite the fact Big Five based on lexiconapproach while FFM based on factor analysis of questionnaire results [71].The Big Five is a trait theory that model five independent domain traitsnamely Openness to experience, Conscientiousness, Extraversion, Agree-ableness and Neuroticism.Openness to experience: Refers to the extent, people choose novelty

12

over the convention, how intellectually curious and creative. High open-ness can be perceived as intellectually curious, individualistic, imaginativeand unconventional whereas sometimes unpredictable or lack of focus. Lowopenness can be characterized as traditional and conservative while seek togain fulfillment through perseverance, and are characterized as pragmaticand sometimes even perceived to be closed-minded.Conscientiousness: Refers as a preference to be organized and depend-able, self-discipline, acting dutifully, planning and aim to achieve ratherthan behaving spontaneously. A high score on Conscientiousness can be de-scribed as reliable, organized, perfectionists and efficient whereas low scoreon this trait characterized as spontaneous, careless, absent-minded and dis-organized.Extraversions: Refers to seek stimulation in the company of others whichis being comfortable with the external world rather than the one’s own com-pany. High extraversions are described as energetic, active, talkative, social,outgoing and sometimes as domineering and attention seeking. The peoplewho are scoring low on this trait are perceived as reserved, quiet, or with-drawn.Agreeableness: Refers to be corporate and compassionate over suspiciousand antagonistic. High agreeableness is perceived as trusting, soft-hearted,generous and sympathetic while it is perceived as naive or submissive. Thosewho score low on this trait is to be competitive, self-confident or aggressive,stubborn and untrustworthy.Neuroticism: Refers to experience unpleasant or negative emotions easilywhich refers to as degree of stability and impulsive control. High Neuroti-cism is to be perceived as being anxious, nervous, moody and worrying incontrary scoring low is on this trait is characterized as emotionally stable,optimistic, and self-confident [44, 98].

It is difficult to characterize individuals as good or bad personalitiesowned, relying on these traits but scoring high or low in each trait has itsown advantages and disadvantages.

2.5.2 Personality in Recommendation Systems

The users who are having different personalities prefer items with differentfeatures which becomes valuable information in the personalized recommen-dation and many researchers have considered personality based recommen-dation in different aspects as

� personalize of items diversity

� addressing the cold start and data sparsity problems

� increase the user satisfaction, etc.[51]

13

Below table describes few personality based recommendation systemsresearch works.

Table 2.3: Personality based Recommendation

Author/s PersonalityModel

Research Direction inRecommendation

Contribution to theDomain

Mehdi et al.[36]

FFM Cold-start Problem inCollaborative Filtering

User-user similaritymeasurement based onpersonality

Matthias et al.[17]

Big Five Cold-start Problem inCollaborative Filtering

User-user similaritymeasurement based onpersonality

Wen Wu etal./Tintarev etal. [101, 93]

FFM Cold-start Problem inCollaborative Filtering

Active Learning basedon personality basedbinary prediction

Lara et al. [79] Big Five Diversity Personality-based di-versity adjusting ap-proach for movie rec-ommendation

Kompan et al.[60]

FFM Diversity Diversity adaptationbased on PersonalityTrait, Openness toexperience

Roshchina etal. [83]

Thomas-KilmannConflict Model

Group Recommenda-tion

Combining assertive-ness and cooperative-ness into the aggrega-tion function

Park [75] Thomas-KilmannConflict Modeland Big Five

Group Recommenda-tion

Group satisfactionmodeling with a per-sonality based graphmodel

Hu et al. [50] Big Five User Satisfaction By personality-basedprofiles based user re-views

Jianwang et al.[100]

Big Five Data Sparsity andScalability

based on the selec-tion of optimal per-sonal propensity vari-ables in collaboratingfiltering

14

Braunhofer etal. [16]

Big Five User Modeling User modeling basedon personality for theusers who lack of do-main knowledge

Yik et al. [102] Big Five Cold-start Problem inCollaborative Filtering

Combine rating fromitem-personality andthat from traditionaluser-based collab-orative filtering inorder to obtain item-personality-basedcollaborative filtering

Johnson, JohnA [56]

n/a Cold-start Problem inContext-aware Recom-mendation

Proposes ademographics-basedContext-Aware Ma-trix Factorizationfor item categories(CAMF-CC) variantthat addresses thisproblem by profilingusers through knownusers attributes, suchas gender, age group,personality traits

2.5.3 Personality Implication on Emotion

Many theoretical frameworks have been able to link users enduring individ-ual traits and transient affective states. In the Yik et.als study, they havefound that all the five personality traits of FFM influence human feelingsand emotional behavior with the evidence that E and N are super factorsfor the affect behavior while O, C and A is in a nonzero Correlation [102].Johnson provided a notable explanation on how personality traits implicateon emotion by each individual traits values in high or low as described inTable 2.4 [56] and their study also traced the at E and N are super factorsfor human emotions.

15

Table 2.4: Personality Implication on Emotion by Jonson (2009)

High Low

Extrovert tend to react with

enthusiasm and of-

ten have positive

emotions

be quiet, low-key and

disengaged

Neuroticism values are emotionally

reactive * in a bad

mood which strongly

aects their thinking and

decision making

calm, emotionally

stable and free from

persistent bad mood

Openness individualistic, non-

forming and are very

aware of their feelings

simple and straightfor-

ward thinking over com-

plex, ambiguous and

subtle

Conscientiousness tend to be prudent tend to be impulsive

Agreeableness reflects cooperation and social harmony

2.6 Affect and Personality Acquisition and

Context-aware Datasets

2.6.1 Context-aware datasets with affect and personality

In any of the research, availability of appropriate data set is a must andmost of the researchers in context-aware recommendation domain face thechallenge of lack of properly recorded dataset with psychological attributes.Since these data is rare most of the researchers tend to obtain their owndatasets for the research works. Following table (Table 2.5) shows few datacorpses, found in the literature for the recommendation systems.

16

Table 2.5: Personality and Emotion based Data Corpuses [95, 26]

Dataset Domain Personality Emotion Users Items

LDOS-

CoMoDa

Movies Big five Six basic

emotions

235 1300

LDOS-

PerAff-1

Images Big five VAD

Space

52 70

LJ2M Blogs Blogs 132

mood

tags

649,712 1,928,868

DEAP Music

Videos

No VAD

Space

32 120

myPersonality Social Net-

work

Big five No varies varies

1000 songs Music No VA space 100 744

ANET Text No VAD

Space

n/a n/a

IDAS Sounds No VAD

Space

n/a n/a

ANEW Text No VAD

Space

n/a n/a

IAPS Images No VAD

Space

n/a n/a

Chittaranjan Mobile

phone usage

Big five n/a n/a n/a

*n/a: not available

2.6.2 Acquisition Methods

Emotion or Affect Acquisition

The affect can be acquired by using explicit or implicit methods. By ask-ing the users directly how they feel or their emotion will be more accurate

17

though it creates a more burdensome experience for users in using recom-mendation service. Therefore, many implicit methods have been introducedand applied in many studies. As examples, affect is acquired using Audiosignals, Language, Visual Signals or Tactile signals where the emotion isrepresented as continues scale or discrete scale [95].

Personality Acquisition

Acquisition of personality has been done mostly by self-reported assessmentby a questionnaire. These questionnaires, consist of five traits descriptiveitems to asset a person. The recent advancements in technologies, newperspectives of acquisition of personality has been arisen by analyzing usercontents from video, audio, mobile and wearable devices, text, social media,etc. [95].

2.7 Recommendation System Techniques

Any of the recommendation system undergo three main phases namely In-formation collection, Learning and Recommendation phases [55]. In thestate of the art of recommendation techniques, three main different cate-gories namely; content-based, collaborative filtering based and hybrid ap-proach are discussed. Apart from these traditional approaches, there arefew recommendation approaches have arisen in modern literature which arecontext-aware recommendation, semantic-based approach, cross-domain ap-proach, peer-to-peer approach, cross-lingual approach etc. [8]. Followingsections discuss few main approaches of above mention especially main tra-ditional approaches and the context-aware recommendation approach whichis the focus of our study.

2.7.1 Recommendation Techniques

Content-based Filtering (CBF)

The content-based recommendation systems, suggest items based on userprofiles that consist of features extracted from the content of the items thatusers already rated before. The recommendations made by this approach areparticularly similar to the items which already have been preferred by theuser. The CBF finds similarity between the items, based on different typesapproaches; Vector space models such as Term Frequency Inverse DocumentFrequency (TF-IDF), Probabilistic models such as Nave Bayes classifier,Decision Trees or Neural Network [38, 33, 13].

For the TF-IDF approach, the weight of a feature k for item i is com-puted as a combination of the Term Frequency (TF) and the Inverse Docu-

18

ment Frequency (IDF):

TFIDF (k, i) = TF (k, i) · IDF (k) (2.1)

The TF factor counts the number of times f is associated with i. In thiscontext, it is usual to normalize the TF values as follows:

TF (k, i) =count(k, i)

maxk’ count(k′, i)(2.2)

The IDF is computed as,

IDF (k) = log|I|nk

(2.3)

where nk is the number of items in item I that consist of feature k.The CBF does not rely on the profiles of the users hence it has the

capacity of adjusting its recommendations quickly and to avoid new userproblem and initial information about users preferred items are gathered atthe initial stage of the user interaction in such recommendation system.

Collaborative Filtering

The collaborative filtering (CF) approach makes the suggestions based theopinions of other users who share similar interests and thus this approachis domain independent. CF techniques create a relation between users anditems which is called use-item matrix for who are having similar tastes toget items as predictions for a certain user for an item which is yet not ratedby him or her. The CF techniques can be divided into two types as memory-based or model-based.

Memory-based CF

The memory-based approaches are based on neighborhood who are havingsimilar tastes in products to specific user and which is further divided intotwo categories as user-based and item-based. The user-based CF approachrelies on the users who share similar preferences of items to a certain userwhile item-based CF recommends items that are similar to the items theypreferred in the past. The CF recommendation process comprise of fourstages mainly; building the user-item matrix, similarity calculation, predic-tive rating calculation and suggesting the top-n items to the users.

Similarity Calculation

The similarity calculation between users or items in recommendationsystem can be done based on various similarity calculation measures like

19

Correlation similarity, Cosine-based similarity, Log likelihood similarity,Euclidean Distance similarity calculation etc. The item-based CF usesthe users who have rated items i and j to calculate the similarity wi,j

between the two items co-rated by users. The user-based CF defends onthe calculated similarity wu,v, between user u and v who have rated thesame items [90].

Correlation-based SimilarityCorrelation-based similarities measures the similarity between two users

u and v Wu,v or the similarity between two items i and j Wi,j. The most pop-ular correlation-based similarity is Pearson Correlation Coefficient which isused to measure into which extent, two variables are linearly related.ThePearson Correlation Coefficient between two users u and v in user-basedalgorithm is stated as

W u,v =

∑i∈I(r(u,i) − ru)(r(v,i) − rv )√∑

i∈I(r(u,i) − ru)2√∑

i∈I(r(v,i) − rv )2(2.4)

where the i ∈ I provides the summation over the items that both theusers u and v have rated and ru is the average rating of the co-rated itemsof the uth user.

The Pearson Correlation Coefficient for the set of users u∈ U who ratedboth items i and j in item-based algorithms is stated as

W i,j =

∑u∈U (r(u,i) − ri)(r(u,i) − rj )√∑

u∈U (r(u,i) − ri)2√∑

u∈U (r(u,j ) − rj )2(2.5)

where ru,i and ru,j are the rating for the user u on item i and j , ri andrj are the average ratings of the ith and jth item by users.

Vector Cosine-based SimilarityThe Cosine-based Similarity is a vector-space model which measures the

similarity between two n-dimensional vectors based on the angle betweenthem. The similarity between two items i and j can be described as W i,j

W i,j = cos(−→i ,−→j ) =

−→i .−→j

|−→i | ∗ |−→j |(2.6)

Log likelihood SimilarityA likelihood ratio test is used to express how many times more likely the

data are under one model than the other. This likelihood ratio can then beused to compute a p-value or compared to a critical value to decide whetherto reject the null model in favor of the alternative model.

D = −2 lnliklihood for null model

likilihood for alternative model(2.7)

20

= −2 ln(liklihood for null model)+2 ln(likilihood for alternative model)(2.8)

The probability distribution of the test statistic of the logarithm likeli-hood ratio can be approximated by using Wilk’s theorem by assuming thenull model is true [9].

Euclidean Distance SimilarityThe Euclidean Distance Similarity is derived as the distance between a

certain user and other each users based on the user-item matrix for eachitem.The mathematical definition of Euclidean Distance Measure is given as

d(xi, yi) =

√∑i=1

(xi − yi)2 (2.9)

This formula states the distance between two users as Euclidean Dis-tance, d and smaller the value, the users are more similar [9].

Predictive Rating Calculation

In CF, obtaining the predictions for a user is a important task and in user-based CF, for an active user a on a certain item i, the weighted averagerating of all the ratings on that item can be used to calculate the predictiverating as follow;

P a,i = ra +

∑u∈U (r(u,i) − ru) ·W a,u∑

u∈U |W a,u |(2.10)

where ra and ru are the average ratings for the user a and user u on allother rated items, and wa,u is the weight between the user a and user u.The summations are concerned over all the users u ∈ U who have rated theitem i.

In item-based CF prediction calculation, the simple weighted average canbe used to predict the rating Pu,i for user u for item i.

P u,i =

∑n∈N ru,nW i,n∑n∈N |W i,n |

(2.11)

where the summation concerned over all other rated items n ∈ N foruser u, W i,n is the weight between items i and n, ru,n is the rating for useru on item n [90].

Model-based CF

The model-based approach always relies on machine learning or data miningtechniques grounded models by employing the users previous ratings to the

21

items. These models mostly extract the features of Clustering techniques,Association Techniques, Neural Networks, Bayesian Networks etc. [55]. Thepredictive rating calculation is one of the main step is in CF: as an example,if its a neighborhood-based CF approach a subset of nearest neighbors ofthe active user is selected while using a weighted aggregate of their ratings isused to predictions based on the similarity with the active user and neighbors[90]. As the final step, the Top-N recommendation derives as a set of N top-ranked items.

Hybrid Approach

Hybrid filtering approaches are out performing solutions for the limitationsof the pure recommendation techniques. Hybrid approaches are mostly acombination of two more techniques for the effective and accurate recommen-dation over pure recommendation techniques.Combination of approachescan be following [3].

� Isolated implementation of algorithms and joining the results

� Exploit some rules of content-based filtering in a collaborative ap-proach

� Apply some rules of collaborative filtering in content-based approach

� Create a unified recommender system, that brings together both ap-proaches

Context-aware Recommendation Algorithms

The context-aware recommendation algorithms mainly divided into two cat-egories as Contextual Filtering and Context Modeling as the context is beingincorporated in the process. The contextual information comprises of situ-ational details of users who experience the recommendation service whichplays an important role in accurate and efficient predictions of items. Thestate-of-the-art of context-aware recommendation algorithms discuss threedetailed ways as described in following to build algorithms [2].Contextual pre-filtering sort out user-item interaction in the recommen-dation that match with the target users context, and use in the tradition2D recommendation algorithms.Contextual post-filtering adjusts the resulting recommendation of anystandard 2D recommendation techniques only based on the contextual in-formation of the target user.Contextual modeling directly exploits the context information within therelevance prediction process. The recommendation algorithms thus, consider

22

multi-dimensional recommendations based on various heuristic and predic-tive modeling techniques. Contextual modeling can be further divided intotwo categories Independent and Dependent contextual modeling. In theindependent approach Tensor Factorization is employed in the recommen-dation. The Dependent contextual modeling either can be Deviation-based:rating deviation between two contexts or Similarity-based: similarity of rat-ing behaviors in two contexts (see figure 2.4) [108, 104].

Figure 2.4: Contextual Modeling Approaches

2.8 User Profiling in Recommendation Systems

A user profile is a set of information extracted to describe a user.Thus,user profiling is the process of obtaining values of different features thatdescribe a user behavior. The RS can be described as direct beneficiary ofuser profiling.The emerging popularity of social network has created newpathways for RSs to build user profiles based on the available sentiments instates, tweets or reviews etc.This leads to build more dynamic user profilesdescriptions rather than the static user profiles which includes explicitlyfilling information from the user through the use of online forms and surveys.

2.9 Evaluation of Recommendation Systems

The evaluation measures for a recommender systems are separated into threecategories namely,

� Predictive accuracy Measures

� Classification Accuracy Measures

� Rank Accuracy Measures

23

Predictive Accuracy Measures

The Predictive Accuracy Measures are used to evaluate how close the rec-ommender system’s predictions of ratings for users.

Mean Absolute Error The most widely used predictive accuracy met-ric is the Mean Absolute Error, or MAE. The MAE calculates the summa-tion of the difference between the user’s rating and the predicted rating anddivide it by the number of items considered.

MAE =1

|Bi |∑bk∈Bi

|ri(bk )− pi(bk )| (2.12)

where Bi denote the items rated by user ai.Mean Squared Error The Mean squared Error measure or MSE em-

phasizes the large error by squaring each individual error.

MSE =1

|Bi |∑bk∈Bi

(ri(bk )− pi(bk ))2 (2.13)

Normalized Mean Absolute Error The Normalized Mean AbsoluteError or NMAE normalizes the MAE by the range of the available ratingvalues. This normalization is employed in order to allow inner data setcomparisons.

NMAE =1

rmax − rminMAE (2.14)

where rmax is the maximum value of the ratings and rmin is the lowest.The Predictive accuracy measures asses the accuracy of the actual rat-

ings.

Classification Accuracy Measures

The Classification Accuracy Measures consider to what extent the Recom-mendation System is able to correctly classify items as interested to user ornot.

Precision and RecallThe precision and recall are two dominant evaluation metrics in Infor-

mation Retrial field for classification tasks.The Precision is defined as theratio of relevant items selected to the number of items selected while Recallis defined as the ratio between the number of relevant items selected andthe total number of relevant items in set of all items.

Precision =|set of selected items that are relevant|

|set of selected items|(2.15)

24

Recall =|set of selected items that are relevant|

|set of all items|(2.16)

F-measure The Precision and Recall measures are inversely related andthe f-measure takes into account both of these metrics and among the variousmetrics that consider precision and Recall together the F-measure is themost widely used.

F −measure =2 · Precision ·RecallPrecision+Recall

(2.17)

Average Precision The Average Precision at the rank of each relevantitem is calculated as below in the formula.

AP =1

|Brs |∑b∈Brs

Precision(b) (2.18)

where Brs is the set of selected items that are relevant.

Rank Accuracy Measures

The Rank Accuracy Measures evaluate a Recommender system’s ability torecommend ordered list of items to a user by assuming that the order of theitems on the list is important.There are several correlation measures thatcan be used to determine the similarity of two list of recommendations.

Spearman’s ρ correlationThe Spearman’s ρ correlation is widely used rank accuracy metric which

is defined as below.

ρ =(x− x)(y − y)

n · stdev(x) · stedv(y)(2.19)

where x1,.....,xn ∈ X and y1,....,yn ∈ Y are the ranks of the items listedin the user’s ranked list and recommendation list respectively and stdev(x)and stedv(y) are standard deviation of x and y .

2.10 Sentiment Analysis Techniques

Semantic Analysis(SA) is extraction of opinion, sentiments and subjectivityof text computational methods.SA is a classification task which undergoesseveral steps like collection of text contents, sentiment identification, fea-ture selection, sentiment classification and find the polarity.SA mainly canbe as two approaches as Lexicon-based approach and Machine Learning ap-proach. The Machine Learning(ML) approach uses most of the ML algo-rithms and linguistic features whereas the lexicon-approach uses sentiment

25

lexicons. The lexicons approaches are further divided into dictionary-basedand corpus-based approaches which riles on statistical or sentiment methodsin defining semantic polarity [68].

2.11 Closing Discussion

In this chapter, we discuss comprehensive related work on emotion-basedrecommendation and personality recommendation. We discuss the motiva-tion of the study by analyzing the importance of emotion as context whileexploring the existing research work in Travel Recommendation, notionsof the recommendation systems and context-aware recommendation, userprofiling, semantic analysis techniques and etc,.

26

Chapter 3

Deriving a Dataset for a

Context-aware Travel

Recommendation

3.1 Introduction

Context-awareness has been introduced into recommendation systems toensure that both long-term and short-term user needs are recognized byconsidering not only preference history but also the current situation. Thisis because if the recommendation process only relies on preference history, itmay not correctly discard an isolated situation such as a gift purchase (i.e:if a user buys a gift for a friends child; a baby suit, user receives sugges-tions for the baby items repeatedly) because it cannot adapt to the currentsituation. Context-aware Recommendation Systems (CARs) therefore in-corporate contextual information including location, time, and activity, andeven advanced parameters such as emotion and personality.

Many such systems have been developed in domains such as movies,books, and music, and emotion is a contextual parameter that has alreadybeen used in those fields. This paper focuses on the use of emotion as acontextual parameter in a tourist destination recommendation system. Wedeveloped a new corpus that incorporates the emotion parameter by em-ploying semantic analysis techniques for destination recommendation. Wereview the effectiveness of incorporating emotion in a recommendation pro-cess using pre-filtering techniques and show that the use of emotion as acontextual parameter for location recommendation in conjunction with col-laborative filtering increases user satisfaction.

In development of the proposed system, we addressed the paucity of

27

datasets with contextual parameters by employing Sentimental Analysis(SA) to acquire the emotional states of users based on user reviews for100 destinations. We then derived emotional tags and manually com-paredthe accuracy of the tags for each of the destinations.

In this chapter, a CARs that uses partial contextual user preferences inthe form of user, item, context and rating was developed. This is unlike atraditional recommendation system, which is based on knowledge of pref-erences of a set of items and the input data is in the form of user, itemand rating. The contextual information that we consider here can be ap-plied to various stages of the recommendation process. A CAR process,based on contextual user preferences elicitation and estimation can take oneof three forms as: contextual pre-filtering, contextual post-filtering or con-textual modeling. In this study, we implemented a contextual pre-filteringparadigm-based solution and use the contextual information to select themost relevant item x user data for generating recommendations [4]. Thepre-filtering approach allows for the deployments of any of the numeroustraditional recommendation techniques previously proposed [3].

3.2 Research Objectives

The specific objective of this chapter are as follows:

1. Derive suitable datasets for tourist destination recommendations con-sidering emotion as a contextual parameter

2. Examine how emotion influences recommendation of destinations us-ing pre-filtering techniques

3.3 Background and Related Work

The state of the art related to recommendation systems is constantly ex-panding. Many of the previous studies have introduced algorithms for rec-ommendations, improved ways of building user models to represent userpreferences, interests and behaviors, and domain specific applications [42].Inthe previous chapter we discussed several studies on Travel recommendationsystems.

Collaborative filtering is a widely adapted recommendation algorithmin which predictions and recommendations are based on the ratings or thebehavior of the users in the system [81, 40].

The approach considers a list of n number of users U and list of mnumber of items I. As shown in Figure 3.1, the user ua,

ua ∈ U

28

Figure 3.1: Collaborative Filtering Process

is called the active user for whom the task of the collaborative filteringalgorithm is to find an item likeliness that can be in two forms. Pa,j is thepredictive rating for item i for the active user Ua, where

j ∈ m

. The prediction rating is a numerical value with the same scale as the ratingprovided by user ua. The recommendation is the list of N items that theactive user will like the most [85]. Collaborative filtering techniques can beeither user-based or item-based. The user-based collaborative filtering modelrecommends items based on computing similar neighbors and creates a groupof users that are compatible with a target user. Item based collaborativefiltering computes similarity based on items and finds items that are similarto the given users rated items [97]. A traditional recommendation systemstarts by estimating the rating function R using an initialized set of ratingsthat are either explicitly provided by users or is implicitly inferred by thesystem for user u and item i that has not been rated.

R: User x Item �RatingIn a CARs, preferences are predicted by incorporating available contex-

tual information into the recommendation process and R estimated byR: User x Item x Context �Rating

‘where User and Item are the space of users and items and Rating isthe space of rating for the user and item pairs. The contextual informationcan be applied during various stages of the recommendation process and theform of the context-aware recommendation can be contextual pre-filtering(contextualization of input), contextual post-filtering (contextualization ofoutput) or contextual modeling (contextualization of recommendation func-tion) [4]. In contextual pre-filtering, the contextual information is used toselect the relevant set of ratings, and ratings are predicted for active usersusing any traditional recommendation function. The context is then set asa query for filtering relevant ratings (see Figure 3.2). The most appropriateuse of pre-filtering technique can vary with the application. One approach

29

Figure 3.2: Pre-filtering Process

is to use a model that targets a local context model for each situation. An-other is to use generalized pre-filtering, which allows for the generalizationof the data filtering query based on a specific context [2]. Generalized pre-filtering states the ratings based on the related contextual situations andderives a collection of prediction models based upon the ratings for eachsegment. Affective computing and SA are combined in various researchfields and many of advanced SA techniques have been developed and manycommercial and academic tools, such as IBM1,SAS2, Oracle3,SenticNet4,andLuminos5 emerged for facilitating polarity evaluations and or mood classifi-cation though most of them are costly and highly limited to set of emotions[20]. SA is considered as computational treatment of opinions, sentimentsand subjectivity of texts while opinion mining is a tedious task since a com-prehensive knowledge of most of the explicit and implicit, regular and irreg-ular, syntactic and semantic rules of a language [68, 21]. In the state of theart of Sentiment classification techniques, the lexicon approaches relay ona sentiment lexicon and the classifications methods using machine learningcan be further divided into supervised and unsupervised learning methods.The supervised learning methods comprised of many classifiers includingProbabilistic classifier, Navie Bayes classifier, Linear classifier, Support Vec-tor Machine classifiers, Neural Network, Decision tree classifiers, Rule-basedclassifiers and Meta-based classifiers [68]. Currently the SA techniques areenriched with deep learning approaches like Deep Convolutional Neural Net-works as well. The Convolutional Neural Networks(CNN) are very similar toordinary Neural Networks (NN), the main difference is the number of layerswhere CNN are just several layers of convolutions with nonlinear activationfunctions applied to the results while traditional NN each input neuron isconnected to each output neuron in the next layer. In CNNs, instead, con-volutions are used over the input layer to compute the output. This resultsin local connections, where each region of the input is connected to a neuronin the output Each layer applies different filters, typically hundreds or thou-sands and combines their results [64, 7]. A drawback of CNN as a classifier

1www.ibm.com/analytics2www.sas.com/social3www.oracle.com/social4www.business.sentic.net5www.luminoso.com

30

is that it finds only a local optimum, since it uses the same back propagationtechnique as MLP [78].

3.4 Incorporating Emotion into Place Recommen-

dation

3.4.1 Emotion tag acquisition

The main challenge with a CARs domain is the lack of appropriate contex-tual datasets. The LDOS-PerAff-1 corpus is one of the datasets created tofulfill this issue [105]. It incorporates video clips of users responding to emo-tional stimuli and ratings with personality traits. Data acquisition was per-formed by presenting a set of images and asking subjects to rate the imagesas if they were choosing images for their computer wallpaper. The LDOS-CoMoDo corpus is another dataset introduced as a context-aware moviedataset comprising of 12 contextual dimensions with 2291 ratings rated by121 users on 1232 items. Among the contextual parameters suggested, threeemotional dimensions were included; endEmo: emotional stated at the endof the movie, domEmo: the emotional state experienced most during themovie and mood: the emotion of the user, when they are watching themovie [70]. The acquisition of emotion has been investigated using a varietyof technologies such as the detection of facial expressions, emotion inferencefrom sensors and other approaches based on voice, speech, body languageand postures. The difficulty in applying these techniques in the recommen-dation system domain is the complexity in adapting these techniques to thesystem implementation. Therefore, previous studies have focused on emo-tion states inferred from reviews by using SA theory. SA techniques can beused to extract emotion from review texts including joy, sadness, fear, anger,and surprise [68]. Emotion Detection can be implemented using MachineLearning or Lexicon-based approaches, with the latter more frequently used[70].

In this study, a context-aware dataset was derived by collecting data forthe global top 100 tourist attractions in 2016. The data including descrip-tion, location and images was obtained from Wikipedia6 , while the averagerating and 100 user reviews for each destination were collected from Tri-pAdvisor7 . The reviews were analyzed and classified to acquire emotiontags to represent user emotional states for a place in two stages. First, weused ToneAnalyser8 , which measures emotional tone, to get a sense of theoverall tone of the review (joy, fear, sadness, disgust, and anger). Second,

6https://en.wikipedia.org/7https://www.tripadvisor.com/8https://tone-analyzer-demo.mybluemix.net/

31

we expanded the positive emotion scale from ’joy’ to ’joy, anticipation, trustand surprise’, and employed SA techniques. Text pre-processing techniqueswere used to derive emotion tags from reviews for each location. A listof words indicative of each emotion was used. The emotion lexicon fromthe National Research Council Canada (NRC), which is based on Plutchikseight emotions and two negative and positive sentiments [53] was used tocalculate term frequencies.

The lexicon set derived by NRC (see Table 3.5) comprises of unigramsand bigrams of the Macquarie Thesaurus [12], all terms in the General In-quirer and the WordNet Affect Lexicon. Thus, the used lexicon set com-prises of unigrams and bigrams of the Macquarie Thesaurus:800 and 787lexicons as adjectives, adverbs, nouns and verbs respectively, all the termsin the General Inquirer: 8132 lexicons as negative, positive and neutral andthe WordNet Affect Lexicon:640 lexicons representing emotion categoriesAnger, Disgust, Fear, Joy, Surprise and Sadness(see Table 3.1,3.2 and 3.3).

Table 3.1: Unigrams/Bigrams from Macquarie Thesaurus.

EmoLex-Uni EmoLex-Bi

Adjectives 200 200

Adverbs 200 187

Nouns 200 200

Verbs 200 200

Total 800 787

Table 3.2: Wordnet Affect Lexicon.

EmoLex-Gi

Anger 165

Disgust 35

Fear 100

Joy 165

Surprise 120

Sadness 53

Total 640

32

Table 3.3: Terms from General Inquirer.

EmoLex-Gi

Negative 2119

Neutral 4226

Positive 1787

Total 8132

In the ToneAnalyzer analysis, for each emotion, a score of less than 0.5indicates that the emotion is unlikely to be perceived in the content and ascore greater than 0.75 indicates a high likelihood that the emotion will beperceived. The overall results show that the highest tone value was recordedin the joy group for all locations, and 88% of the joy group indicates thatthe emotion is likely to be perceived in the selected review texts (see Table3.4).

Table 3.4: Sample results of tone analysis created by ToneAnalyzer

Place ID Anger Disgust Fear Joy Sadness

10 0.0734 0.1146 0.1115 0.5307 0.2341

11 0.0744 0.1156 0.1098 0.5013 0.2595

12 0.0824 0.0660 0.0635 0.6170 0.2659

13 0.0913 0.1158 0.1204 0.5066 0.2320

14 0.1519 0.0763 0.0700 0.5670 0.2680

15 0.1353 0.0644 0.1011 0.5452 0.2609

16 0.1426 0.1612 0.1651 0.4229 0.3192

17 0.0951 0.0613 0.1521 0.5133 0.2649

18 0.1232 0.1023 0.1132 0.5245 0.3603

19 0.0928 0.0865 0.1135 0.5266 0.2406

20 0.0673 0.0728 0.0742 0.6117 0.2855

Text pre-processing is a significant task in text mining techniques, andits application is the first step in any system. The main aim behind it is torepresent each document as a feature vector, so it separates the text into in-dividual words. The quality of the classification process is highly dependenton this feature selection process. Therefore, it is important to select mean-ingful key-words and discard words that do not enable the distinguishing of

33

the documents [6].

Table 3.5: Extract from the emotion lexicon used in the classification

Joy absolution abundance abundant accolade

Surprise abandonment abduction abrupt accident

Anticipation abundance accelerate accolade accompaniment

Trust abacus abbot absolution abundance

Fear abandon abandoned abandonment abduction

Anger abandoned abandonment abhor abhorrent

Sadness abandon abandoned abandonment abduction

Disgust aberration abhor abhorrent abject

Next, we performed a Term Frequencies (TFs) calculation based on thelexicons of the eight emotion groups. We chose the emotion category fora destination based upon the maximum frequency value as shown in Table3.4. The TFs were calculated for each review and the total TFs calculatedfor each emotion. We then selected the emotion tag for each destinationbased on the highest frequency (see Table 3.6).

Table 3.6: Emotion tag derivation based on TF values

Place ID User ID An Ant D F J S Su T

1 1 1 3 1 0 4 0 2 3

1 2 0 1 0 0 0 0 0 1

1 3 1 3 1 1 3 2 2 2

1 4 1 0 1 0 1 0 0 1

1 5 1 3 0 2 1 1 1 1

Emotion Word Sum 4 10 3 3 9 3 5 8

1 Emotion Tag Anticipation

An- Angry, Ant- Anticipation, D- Disgust, F- Fear, J- Joy, S-Sadness, Su- Surprise and T- Trust

34

Table 3.7: Emotion tag derivation of TFs

Place ID J Su T Ant D F An S Emotion Tag

1 29 14 55 42 6 12 13 16 Trust

2 28 12 32 23 1 10 5 4 Trust

3 31 14 21 24 0 3 4 10 Joy

4 17 7 17 13 0 6 4 6 Joy

5 39 18 44 44 5 14 10 11 Trust

6 31 32 43 41 8 11 16 18 Trust

7 51 26 41 38 2 3 2 8 Joy

8 46 17 48 39 9 11 10 9 Trust

9 24 13 22 22 2 8 6 11 Joy

10 23 11 28 21 1 6 4 5 Trust

11 45 19 49 36 5 12 4 7 Trust

12 44 31 26 64 4 7 7 5 Anticipation

13 18 7 20 16 1 5 4 4 Trust

14 49 21 47 62 6 14 7 9 Anticipation

15 35 16 37 43 2 21 7 12 Anticipation

An- Angry, Ant- Anticipation, D- Disgust, F- Fear, J- Joy, S-Sadness, Su- Surprise and T- Trust

In the case of the occurrence of multiple emotions, we assumed that if anegative emotion category and a positive emotion category appeared withequal frequency, the negative one was stronger. This assumption is basedon the fact that humans are more honest when stating negative emotions.In terms of the positive emotions occurring with an equal frequency weassumed joy, trust, anticipation and surprise are stated in descending orderby considering the fact that human emotion identification and stating abilityby themselves (see Table 3.7).

We collected 9998 ratings from 8470 users for the 100 selected locationsand derived emotion tags for each location for each user. From this informa-tion, we derived two datasets: place dataset place id, name, description,place category, location, average rating, emotion tag and user dataset userid, place id, user emotion tag, user rating. Figure 3.3 illustrates the emotiontag derivation process.A = total number of cases that the system assigned emotion tags

35

C = number of correct cases based on manual judgment

Precision =C

A(3.1)

We calculated the precision of emotion tag detection and compared theresults manually by reading review texts for the place dataset. The resultsshow that the precision of the detection process is 63.3 % [91].

Figure 3.3: Emotion Tag Creation

3.4.2 Proposed system

We implemented the proposed recommendation system using the two de-rived datasets and loaded the location data into the database of the system.The dataset was input to the recommendation function based on the emo-tion state of the user when they logged in, as shown in Figure 3.4. Usingthe pre-filtering techniques in CARs, the similarity and predictive ratingvalues were calculated for each user and the top five place recommendationswere generated. The systems recommendation engine consists of two phasesbased on collaborative filtering without emotion (CFN) and collaborativefiltering with emotion (CFE). In the implementation, we used item-item

Figure 3.4: Emotion Gathering

collaborative filtering to develop and review our contextual parameters onthe derived dataset. The recommendation process is illustrated in Figure3.5. Each recommendation based on emotion was analyzed by consideringthe three emotion groups derived according to the Plutchik emotion classi-fication system. In the recommendation process, data was selected for the

36

Figure 3.5: Recommendation Process

location recommendation engine based on these three groups from the as-sumption that the recommendation should fall on the positive emotion scale.The results of the ToneAnalyzer analysis reinforced the use of the positiveemotion scale because the collected reviews were more likely to have a posi-tive emotion. Therefore, disgust, fear, anger, and sadness were re-arrangedinto three positive groups (anticipation, joy and trust) based on Plutchikscomprehensive list of eight primary emotions arranged as opposing pairs (seeTable 3.8). To avoid a negative emotion category, fear and anger were allo-cated to the joy and trust groups, respectively. Further, we used these threegroups to evaluate the influence of user emotion on the recommendation.

Table 3.8: Sample results of tone analysis created by ToneAnalyzer

Group I Anticipation Anticipation, Surprise

Group II Joy Joy, Sadness, Fear

Group III Trust Trust, Disgust, Anger

In the recommendation process, a prediction of a target users rating onan unrated target item was calculated by considering the users rating ofobserved items. This allows for user-item rating pairs to be used to ratevalue predictions, as shown in Figure 3.6 [41].

For item-item collaborative filtering, users who have rated both item iand item j are identified and then the similarities are computed [90]. Thesimilarity calculation is performed based on measures such as the Pearsoncorrelation, Euclidean Distance, Tanimoto Coefficient or the LogLikelihoodSimilarity. In the proposed process, the similarity calculation was basedon the Loglikelihood ratio, which relies on the statistical similarity betweentwo items or users and which yielded a sufficient number of items for therecommendation. The Loglikelihood ratio utilizes occurrences related to

37

Figure 3.6: User-item Matrix

users or items such as users or items that overlap and the events for whichboth users or items do not have preferences [34, 49]. Prediction algorithmsestimate the rating that a user would provide for a target item [74]. Foritem-based prediction, the Simple weighted average can be used to predictthe ratings [34]. Here, we calculated the predictive rating Pu,i by user u foritem i as follows:

Pu,i =

∑n=N (sim(u,n) + 1)×Ru,n∑

n=N (sim(u,n) + 1)(3.2)

where sim(u, n) is the similarity between the nth item and user u, andRu,n is the rating by user u of item n for all N number of items that are basedon the Mahout item-based recommendation algorithm [86]. The similaritycalculations ranged from 1.0 to 1.0, and to avoid negative values, we added1.0 to similarity values so that the similarity ranges from 0.0 to 2.0. The topfive place recommendation list was created based on the highest similarityvalues from the most similar places set from the places pool.

3.4.3 Predictive rating calculation (item-item collaboration)

The predictive rating calculation is illustrated below by using an exampleof user-item matrix and similarity values in the recommendation process.

Pu1,s1 =(1 + 1) ∗ 1 + (1 + 0.75) ∗ 0 + (1 + 0.66) ∗ 1 + (1 + 0.8) ∗ 1

(1 + 0.75 + 0.66 + 0.8)(3.3)

Pu1,s2 =(1 + 0.75) ∗ 1 + (1 + 1) ∗ 1 + (1 + 0.5) ∗ 0 + (1 + 0.6) ∗ 1

(0.75 + 1 + 0.5 + 0.6)(3.4)

38

Figure 3.7: Sample User Behavior Actions

Figure 3.8: Loglikelihood Similarity Values

where Pu1,s1 is predictive rating for user u1 for place s1 and Pu1,s2 is pre-dictive rating for user u1 for place s2.

Figure 3.9 shows an example of top five place list provided for a user.

Figure 3.9: Recommendation List

3.5 Experimental setup and evaluation

Experimental setup

The Travel Destination location recommendation system was presented to16 users. Each user was asked to evaluate two recommended lists accordingto the user preference for each location and the overall preference for the

39

list according to the users current emotion and overall satisfaction basedon the five point-Likert Scale. The evaluation was performed to assess theusers opinion of the quality of the CFE recommendation algorithm so thata baseline comparison with the CFN algorithm could be performed.

Evaluation

We used Precision and Mean Average Precision (MAP) values of the twoapproaches in our evaluation.

As well, a t-test analysis was performed to test the superiority of CFEagainst the base-line approach. The t-test evaluates the tmean of both Aver-age Precision (AveP) Values and Average Preference Ratings (APR) basedon Preferred and Preferred much user ratings in the five point Linkert scale.

Moreover, we evaluated the recommendation list by considering the emo-tion groups derived at the recommendation engine design stage to track howthe lists correspond with user emotions.

The overall user satisfaction of the recommendation system was alsoanalyzed. Users were asked to input their emotion from the emoticon scaleand evaluate the two lists of five locations.

3.6 Results

We used the Precision, Classification Accuracy Measure in our evaluation.This requires a binary Do not recommend/select, Recommend/select scale,so we assumed that ratings of 4 and 5 were good recommendations [88].

Table 3.9: Precision and Recall matrix

Selected Not Selected Total

Relevant Nrs Nrn Nr

Not Relevant Nis Nin Ni

Total Ns Nn N

Based on the Precision Recall Matrix (see Table 3.9) precision is statedas;

Precision =|N rs||N s|

(3.5)

The precision values for the CFN and CFE for the 16 users were cal-culated as below and the mean precision values for the CFE was greatercompared with CFN. Average Precision calculates the precision at the posi-tion of every correct item in the ranked results list of the recommender. The

40

mean of these average precision values across all relevant lists is the meanaverage precision or MAP. The MAP is also greater for CFE compared toCFN (see Table 3.10).

AveP =

∑nk=1(p(k)× rel(k))

number of relevant items(3.6)

MAP =

∑nk=1AveP (q)

Q(3.7)

where P(k) is the precision at k-th element, rel(k) is 1 if the i-th item ofthe list is relevant and Q is the total no. of lists.

Table 3.10: Precision and Mean Average Precision values

Algorithm Precision (%) Mean Av. Precision (%)

CFN 59.69 64.4

CFE 65.31 73.8

Moreover, we analyzed the Mean Average Precision based on emotionalgroups (MAPE) for each approach:

MAPE =

∑Cc=1

∑qq=1AveP (q)∑Cc=1Q

(3.8)

where C is the no. of emotion groups based upon three groups.

Figure 3.10: Mean Average Precision Values with Emotion Groups

In Figure 3.10, denote the collaborative filtering approach for the trust,joy and anticipation emotional groups, respectively.

We compared the performance of the both CF approaches, in terms ofthe mean of average precision values and the mean of average preferenceratings.

41

Mean Of Average Precision Comparison for CFE and CFN:Hypothesis are:

H0: µc= µCFN and alternatively Ha: µc 6= µCFN, Ha: µc > µCFN,where µc and µCFN are the mean average precision rating of the context

aware and non-context collaborative filtering approaches, respectively.

Since T (Test Statistic) > tα ,γ(Critical Value), we reject the nullhypothesis and population means are different at the 0.05 significancelevel while for the alternative hypothesis µc > µCFN and µcp > µCFN.Therefore, the test results show strong statistical significance with thebaseline recommender (CFN)(p-value = 0.044).

Mean Of Average Preference Comparison for CFE and CFN:Hypothesis are:

H0: µcp= µCFN and alternatively Ha: µcp 6= µCFN, Ha: µcp > µCFN

where µcp and µCFN are the mean preference ratings of the contextaware and non-context collaborative filtering approaches, respectively.

Since T (Test Statistic) < tα ,γ(Critical Value), we don’t reject thenull hypothesis in both cases and conclude that the population meansare closely equal at the 0.05 significance level. Therefore, the test resultsshow that the difference with the baseline recommender (CFN) in terms ofaverage preference ratings (p-value = 0.285) of CFE was not statisticallysignificant.

Finally, user feedback on overall satisfaction with the recommended listas well their opinion based on their current emotion is shown in Table 3.11.According to the results, 60% of users were satisfied overall with the recom-mended lists. Further, 53% acknowledged that the provided list matchedtheir current emotion.

Table 3.11: User satisfaction

Algorithm Overall preference (%) Preference with emotion (%)

CFN 60 53.33

CFE 46.67 -

3.7 Conclusion

In this study, we established how emotion can impact the travel destinationrecommendation process. The use of emotion as a contextual parameter

42

for location recommendation in conjunction with collaborative filtering in-creased user satisfaction. In addition, we derived emotion tags for each lo-cation based on user review to examine how the destination can be effectedby emotion in a travel destination recommendation system. While previousstudies have incorporated emotion into recommendations for predefined in-door places, our study incorporated it onto a recommendation system forfamous tourist attractions. The accuracy of detecting the correct emotiontag using the lexicon-based approach was 63%. However, we believe thatthis can be improved using other SA approaches.

Plutchiks emotion categorization was used to derive both emotion tagsand the acquisition of the emotions of users, and the recommendation listincorporated positive emotion categories. Moreover, the sensitivity of theemotion contextual parameter in the recommendation was analyzed basedon the accuracy of the lists for the user.

3.8 Closing Discussion

In our approach, we derived the tags based on Plutchik emotion categories.The lexicon-based methods for SA are robust, result in good cross-domainperformance, and can be easily enhanced with multiple sources of knowledgecompared to other approaches.

In deriving emotion tags to enhance accuracy, opinion mining and othersemantic analysis can also be used. The lexicon-based methods dependon the lexicon we are using and few lexicons are appeared in SA researchworks and one such a lexical resource, SentiWordNet which is one dictio-nary of opinionated terms. As well, the deep learning approaches and opin-ing mining techniques explained in the background section can also be usedto enhance the accuracy of emotion word classification. The senticword isbuilt based on SentiWordNet lexicon and adapt Hourglass of emotion. Inthis model, sentiments are reorganized around four independent dimensionswhose different levels of activation, in fact in this model affective states arenot classified into traditional emotional categories rather into four concomi-tant but independent dimensions - Pleasantness, Attention, Sensitivity andAptitude. Although we used exact Pre-filtering, which for the use of tradi-tional recommendation algorithms does not consider any rating acquired insituations even slightly different from the targeted one.

43

Chapter 4

An Analysis of Emotion and

User Behavior for

Context-aware

Recommendation Systems

using Pre-filtering and

Tensor Factorization

Techniques

4.1 Introduction

Currently, information systems are mainly confronted by the challenge ofinformation overload and alleviative methods are required to cope with andultimately overcome this problem effectively and efficiently. Recommen-dation systems act as intelligent agents that provide solutions; the actualrecommendation procedure is stated as ”the process of utilizing the opinionsof a community of customers to help individuals in that community for moreeffectively identifying the content of interest from a potentially overwhelm-ing set of choices.” [81]. The issue of context-awareness has been introducedto the recommendation domain to increase the efficiency and usability of

44

information filtering systems while providing solutions to information over-load. Context-awareness emerged to acquaint users with the influence ofthe external environment on his/her appreciation of the items. Recently,however, not only the physical conditions but also users’ psychological con-ditions are considered in the recommendation process. In the context-awarerecommendation domain, many contextual features have been identified ascontextual parameters, for instance companion in the movie recommenda-tion domain, time and mood parameters in the music recommendation, andweather, season, travel type, etc., in the travel recommendation domain [1].

Much research has examined the application of context on the effectiverecommendation solutions in a variety of domains [10, 32, 14, 63]. In ourstudy, we mainly focus the incorporation of emotion and user behavior ascontextual parameters and try to compare the feasibility of the selectedparameters on one of the challenging domain, travel recommendation com-pared to classical recommendation domains like movies, music, books, etc.Emotion has been used as a contextual parameter in several studies re-cently and a few studies have made the effort to discuss the effectivenessof emotion in the recommendation process [109] with a real world dataset,for example LDOS-CoMoDa which is a movie dataset. Recently, recom-mendation systems have revealed a great tendency to adapt emotion due toits effectiveness in human decision-making processes. These research effortshave been conducted independently and extended to two major research do-mains, i.e. Recommender systems and Affective computing. In particular,the user behavior actions; rating and tagging emotion were added to ourstudy to analyze the effectiveness in the presence of both parameters to-gether and individually in the proposed travel destination recommendationsystem. The context-aware dataset on travel recommendation is difficult tofind compared to traditional recommendation domains. Firstly, we derivedour dataset based on user reviews for each destination chosen for the rec-ommendation by employing Semantic Analysis concepts. Secondly, we com-pared each parameter’s effectiveness on travel destination recommendationby using the Pre-filtering technique and Tensor factorization. Consequently,this study focuses on investigating the context-aware recommendation basedon the parameters, user behavior and emotion. This is done by using thePre-filtering technique in collaborative filtering and Context Modeling inTensor factorization.

The main objective of this paper is to propose a framework for the trav-eler destination recommendation system, exploiting the emotion and userbehavior while optimizing each contextual parameter in the recommenda-tion process by using Pre-filtering techniques and Tensor factorization.

45

4.2 Research Objectives

We propose two major contributions for context-aware travel recommenda-tion. We chose two contextual parameters - emotion and user behavior andwe implement the recommendation by using Pre-filtering techniques andthen compare the effectiveness of the recommendation systems with the ab-sence of any contextual parameter. Furthermore we compare how individualparameters perform when we choose context in the recommendation process.This is followed by implementing the recommendation system with TensorFactorization and comparing the results for each parameter. Finally, wecompare two approaches, specifically Pre-filtering and Tensor Factorizationfor each context. We evaluate all the implemented approaches with preci-sion, mean average precision, mean average precision with emotion groups,the t-test (to check the difference in average precision and average prefer-ence ratings) and overall user satisfaction. The objective is to conclude thefeasibility of selected contexts in travel recommendation.

4.3 Background and Related Work

4.3.1 Context-aware recommendation

Context-aware recommendation systems try to estimate the rating functionR by incorporating the contextual information C for item I by user U withinitially specified user rating as follows;

R: User x Item x Context �Ratingwhere the user× item pair refers to users who are not rated yet. Integrationof emotion in the recommendation process can take three forms, these beingcontextual pre-filtering, contextual post-filtering and contextual modelingaccording to Adomavicius et al. [4]. In the pre-filtering stage the contex-tual information is used to filter and select the most relevant data beforeapplying the recommendation algorithm, while in the contextual-post filter-ing the recommendation algorithm is applied to the original dataset. Thenthe recommendations are filtered according to the contextual information.Both pre-filtering and post-filtering treat the recommendation as a 2D rec-ommender. The contextual modeling approach considers the context in therecommendation algorithm and hence the rating function is treated as a 3Dfunction which represents user, item, and context as User x Item x Contextand provides a rating for each user.

4.3.2 Pre-filtering with context-aware recommendation

The collaborative filtering technique uses the preference for items preferredby users (user-item matrix). Each entry in the user-item matrix represents

46

the preference score or the ratings of the ith user for the jth item, where thereis m no. of users, U = u1, u2, . . ., um and n no. of items, I = i1, i2, . .., in. Each user ui has a list of items ii, that the user has expressed his/herpreference about (see Figure 4.1). By calculating the similarities betweenthe users, collaborative filtering matches no. of users with relevant interestand make the recommendation [46].

Predictions Pi,j, express the predicted user’s preference for item j forthe active user u i. The predictive value is a numerical value which is on thesame scale as the ratings provided by users.

Recommendation List is a list of top N items, that the active userui will prefer most and this list represents a no. of items that user hasn'tpreferred so far. The collaborative filtering can be two categories namelymemory-based (user-based) and model-based (item-based) [18, 61].

Figure 4.1: Collaborative Filtering Process with Pre-filtering Technique.The

filtered user-item matrix data are used to predict rating for a user

In the pre-filtering approach, the user-item matrix is made of the filtereddataset according to the current context in the query.

4.3.3 Contextual modeling with context-aware recommenda-

tion

In applying Tensor Factorization techniques to recommendation systems, apredictive model is provided by analyzing patterns from the data, linkedto the multifaceted nature of user-item interaction. A tensor is consideredto be an array of numbers with more than two dimensions and serve as anatural extension of matrices to a higher order case [27, 43].

Thus, a tensor is a multidimensional array or can refer to an Nth ordertensor which is an element of the tensor product of N vector spaces. Weuse CANDECOMP/PARAFAC(CP) Tensor Factorization model which useratings from M users for N items under Q types of contexts as a three-dimensional tensor M×N×Q. The latent features are stored in UεR M X D,

47

IεR N X D and CεRQ X D where U represents D-dimensional row vector form users, I for latent features for item i and C represents the latent featuresof context category q ; [59] (see Figure 4.2)

fmiq =D∑i=1

UmdIidCqd (4.1)

Figure 4.2: Tensor Factorization Model.

4.4 Proposed System

4.4.1 Context acquisition

In the implementation of the proposed recommendation system, the contex-tual parameters must be acquired and made available for recommendation.

Figure 4.3: Emotion Acquisition.Logged in users’ emotion were collected in

Plutchik’s emotion classification scale.

The emotion of current users is recorded in the system when users arelogged in to it. Users’ current emotions are collected according to wherethey are on the scale for Happy, Surprise, Disgust, Sad, Afraid, Angry,Confidence and Anticipation based on the Plutchik emotion classificationdepicted in Figure 4.3 [77]. With reference to user behavior, we recordeduser actions such as user rating and emotion tagging for places in the systemas illustrated in Figures 4.4 and 4.5.

48

Figure 4.4: User Rating Capturing.

User preference for a place gathered

in five-star scale.

Figure 4.5: Emotion Tag Cre-

ation.The users were asked to tag

an emotion for place when they are

logged into the system.

4.4.2 Recommendation system implementation

The implementation of the proposed recommendation system was donebased on two approaches, Pre-filtering and Tensor factorization. For eachapproach we analyzed the suitability of each contextual parameter emotionand user behavior for the experiment. We developed two recommendationengines for the system which are based on collaborative filtering and TensorFactorization, using the two derived datasets and loaded the place data into the system’s database. The proposed framework for the recommendationprocess with each approach is illustrated in Figure 4.6.

The collaborative filtering approach considers the recommendation pro-cess with emotion (CFE), user behavior (CFUB) and emotion and user be-havior (CFEEUB) together with the Pre-filtering techniques in the proposedcontext-aware recommendation system. The similarity and predictive ratingvalues are calculated for each user and the top five places were recommendedto the user for each CFE,CFUB and CFEUB approaches. To compare thesystems effectiveness, the system which is not incorporated with any contex-tual parameter (CFN) also developed. In the implementation process, weused item-item collaborative filtering to develop and review our contextualparameters on the derived dataset.

In the context-aware recommendation process the recommendationbased on emotion was analyzed with reference to three emotion groups de-rived based on the Plutchik’s emotion classification: Group I: Anticipation(Anticipation, Surprise), Group II: Joy (Joy, Sadness, Fear) and Group III:Trust (Trust, Disgust and Anger). In the collaborative filtering recommen-dation process, selecting data to be placed in the recommendation engine

49

Figure 4.6: Architecture for the Recommendation System with two Rec-

ommendation Approaches. The Tensor factorization follows the CP factor-

ization model while item-item collaborative filtering was used as the Pre-

filtering technique.

was done based on these three groups since we assumed that the recommen-dation should lie on the positive emotion scale. Therefore, Disgust, Fear,Anger, and Sadness are categorized and arranged as three positive groups,i.e. Anticipation, Joy and Trust based on Plutchik’s comprehensive list ofeight primary emotions arranged as opposing pairs. To avoid using the neg-ative emotion category, Fear and Anger were arranged in the Joy and Trustgroups, respectively. We utilized these three groups to evaluate the influ-ence of user emotion on the recommendation. The dataset was input tothe recommendation function based on the user’s emotion states that werelogged in to the system.

Thus, to apply Pre-filtering with item-item collaborative filtering, thetarget user’s rating prediction for a target item is done by considering theuser-item rating pairs, while the prediction for a certain user for a targetitem is predicted based on the user’s ratings on the observed items. Thesimilarity calculation for item i and item j, needs to identify set of users whohave rated for both items and then compute the similarities for those users

50

as illustrated in Figure 4.7 [41, 90].

Figure 4.7: User-item Matrix Illustration

The similarity calculation for the recommendation process was calcu-lated based on the Loglikelihood ratio which relies on calculating the sim-ilarity between two items or users based on statistics since the Loglikeli-hood provided a sufficient number of items for recommendation comparedto Pearson’s Correlation. The Loglikelihood ratio is derived based on theoccurrences relating to the users or items which are users or items overlap-ping in preferences where both compared users have preferences or not, andthe events where both users or items do not have preferences [34, 49].

The prediction algorithms play the role of guessing the rating a userwould provide for a target item [74]. We calculated the predictive ratingPu,i by user u for item i as shown below:

Pu,i =

∑n=N (sim(u, n) + 1)×Ru,n∑

n=N (sim(u, n) + 1)(4.2)

where sim (u,n) is the similarity between nth item and user u and Ru,n isthe rating for user u on item n for all N number of items which are basedon the Mahout item based recommendation algorithm [86]. The similaritycalculations ranged from -1.0 to 1.0 and to avoid negative values we added1.0 to similarity values so the similarity ranges from 0.0 to 2.0. The topfive place recommendations list was created based on the highest similarityvalues to the least in the provided most similar places, set from the placespool.

In the contextual modeling approach, we used Tensor factorization, be-cause it has the advantage of a tensor-based approach with the ability toconsider the multifaceted nature of the user-item interaction. In the CP Ten-sor Factorization, three-dimensional tensors M x N x Q are used for M usersfor N items with the Q types of contexts. The Tensor factorization approachalso followed with two proposed contextual parameters; the recommendationapproaches are implemented as Tensor Factorization with emotion (TFE),

51

Tensor Factorization with user behavior (TFUB) and Tensor Factorizationwith emotion and user behavior (TFEUB). So in each case at least one pa-rameter was selected for the recommendation process. The recommendationof the tensor-based approach was developed based on CARSkit library [107]and the context appearance for user m for item i with the context k can beeither 1 or 0 as a binary value and input to the recommendation process. Asan example, for user 4 for Place ID 52 the contextual information is statedbelow in Table 4.1.

Table 4.1: Example of the Results for Contextual Information for a User

emot

ion

:an

gry

emoti

on

:anti

cip

ati

on

emot

ion

:d

isgu

st

emot

ion

:fe

ar

emot

ion

:jo

y

emot

ion

:sa

dn

ess

emot

ion

:su

rpri

se

emot

ion

:tr

ust

rate

d

tagg

edem

o

TFEUB 4 0 0 0 0 0 0 1 0 1 1

TFUB 5 0 0 0 0 0 0 0 0 1 1

TFE 3 0 1 0 0 0 0 0 0 0 0

Figure 4.8 depicts an example for top five place list provided for a userin Tensor Factorization approaches.

4.5 Experimental Set up and Evaluation Protocols

4.5.1 Experimental setup

First, we describe the evaluation protocols for place recommendation in theproposed system. Then we demonstrate how the contextual parameter se-lection for Pre-filtering and Tensor factorization performed to assess theeffectiveness in the recommendation process. Afterwards we evaluate theimpact of each parameter on each approach in the recommendation processin terms of the recommendation system’s quality. For comparison reasons,we separately evaluate each recommendation approach, and especially thecollaborative filtering approach against non-context collaborative filteringbased recommendation. In addition, for each context, the Tensor Factor-ization and Collaborative Filtering approaches were also compared. Thereason for selecting both Tensor Factorization and Collaborative Filteringis because both filtering and contextual modeling approaches can be tested

52

Figure 4.8: Recommendation Lists. The recommended lists in the figure

show places provided by three recommendation approaches TFE,TFUB and

TFEUB.

in the context-aware recommendation.The implemented system, Travel Destination place recommendation was

presented to 16 users and they were all asked to experiment with the rec-ommendation process and evaluate the two recommended lists. These lists,which were asked for rating according to the user preference for each placein the list and state the overall preference for the list according to user’scurrent emotion while stating the overall satisfaction for recommended listsin the five point-Likert scale.

Evaluation Protocols for the task of recommendation lists, the followingare used:

1. We used Precision and Mean Average Precision (MAP) values of thetwo approaches

2. We evaluated the recommendation lists considering the emotion groupsderived at the recommendation engine designing stage to track, howthe lists are fitted with the users’ emotions by using Mean AveragePrecision on Emotional groups(MAPE)

3. To compare contextual recommendation against non-context rec-ommendation we used t-test to examine the superiority of theall context incorporated approaches (CFE, CFUB, CFEUB, TFE,TFUB,TFEUB) against baseline approach (CFN) by evaluating tmean

of both Average precision (AveP) and Average Preference Rating(APR) based on the rating users marked as Preferred and Preferredmuch in the five point Likert scale

53

4. The user rating behavior variation for the user emotion was analyzedfor the emotion groups validation in the recommendation system de-sign

5. The overall satisfaction of users towards the recommendation systemanalyzed

Thus, in the experiment the testing users had to register to the system andthen input their emotions in the given emoticon scale and evaluate the twolists, each with five places.

4.6 Results

The Classification Accuracy Measure is one of the Accuracy metrics, whichmeasures to what extent a recommendation algorithm can correctly classifyitems as interested or not. In our study, we use Precision one of the popularmeasurers for recommendation system evaluation; to use this, recommendersystem must convert its rating scale into a binary scale so we converted ratingscale as ratings of 4 and 5 are good recommendations [88]. The Precisionexpresses the fraction of recommended items that is actually relevant to theuser [55].

Precision =|Correctly recommended items||Total recommended items|

(4.3)

The precision values for the all the context incorporated approaches (CFE,CFUB, CFEUB, TFE, TFUB, TFEUB) and the non-context approach(CFN) as rated by 16 users, were calculated as below and mean precisionvalues for the all the context incorporated approaches was greater comparedto CFN. Average Precision calculates the precision at the position of everycorrect item in the ranked results list of the recommender. The mean ofthese average precision values across all relevant lists is the mean averageprecision or MAP. The MAP also outperforms for the all the context incor-porated approaches compared to CFN (see Table 3).

AveP =

∑nk=1(p(k)× rel(k))

number of relevant items(4.4)

MAP =

∑nk=1AveP (q)

Q(4.5)

where P(k) is the precision at k-th element, rel(k) is 1 if the i-th item ofthe list is relevant and Q is the total no. of lists.

54

Table 4.2: Precision Values and Mean Average Precision Values

Algorithm Precision (%) Mean Av. Precision (%)

CFN 59.69 61.98

CFUB 61.88 67.28

CFE 65.31 69.83

CFEUB 71.88 79.18

TFUB 63.75 72.27

TFE 68.75 74.81

TFEUB 72.5 81.59

Moreover, we analyzed the Mean Average Precision based on emotionalgroups (MAPE) for each approach.

MAPE =

∑Cc=1

∑qq=1AveP (q)∑Cc=1Q

(4.6)

where C is the no. of emotion groups and we rely on three groups in ourevaluation.

In Figure 9, The overall MAPE values on Collaborative Filtering ap-proach and Tensor Factorization approach with emotional groups Trust, Joyand Anticipation are illustrated respectively. Thus, the emotional group viseMAP showed increased results compared to general approach.

Figure 4.9: Average Precision Values with Emotion Groups

We, statistically compared the performance of the Collaborative Filter-ing approaches and Tensor Factorization approaches with non-context ap-proach CFN; in terms of average precision values and average preference

55

ratings.Mean Of Average Precision Comparison for CFUB,CFE,

CFEUB,TFE,TFUB & TFEUB with CFN:The hypotheses are:H0: µc= µCFN and alternative hypotheses are Ha: µc 6= µCFN, Ha: µc>µCFN.where µc and µCFN are Mean of average precision rating of context awareapproaches and non-context approaches.The test results showed that thedifference with the baseline recommender (CFN)as below.

I TFEUB obtained very strong statistical significance at 95% of signifi-cance level (p=0.0005)

II CFE, CFEUB & TFE approached strong statistical significance(p=0.044, p=0.011 & p=0.038) compared to CFN

III CFUB, approached significance (p=0.067), suggesting moderate corre-lation with CFN

IV TFUB approached significance (p=0.082), suggesting a weak correlationwith CFN

Mean Of Average Preference Ratings Comparison forCFUB,CFE, CFEUB,TFE,TFUB & TFEUB with CFN:

H0: µcp= µCFN and alternative hypotheses are Ha: µcp 6= µCFN,Ha: µcp>µCFN.where µcp and µCFN are Mean of preference ratings of context aware ap-proaches and non-context approaches. The test results showed that thedifference with the baseline recommender (CFN)as below.

I TFEUB & CFEUB obtained, strong statistical significance at 95% ofsignificance level (p=0.019, p=0.012)

II CFE, CFUB, TFE & TFUB the differences were not statistically sig-nificant with CFN

Table 4.3 illustrates p-values obtained for the t-test from MAP and MeanRating Preferences.

56

Table 4.3: Precision Values and Mean Average Precision Values

Algorithm Mean Average Precision(%) Mean Rating Preference (%)

CFUB 0.067 0.376

CFE 0.044 0.285

CFEUB 0.011 0.019

TFUB 0.082 0.131

TFE 0.038 0.147

TFEUB 0.0005 0.012

The user rating behavior showed that users have rated as preferred andmuch preferred when they are Joy, Surprise, Trust and Anticipation emotion,compared to Fear, Sad, Disgust and Angry emotions. Thus, these resultsshowed that suggesting places, incorporated with positive emotion is muchsuitable and the used emotional group classification in the recommendationsystem design process, works fine in this regard.

Figure 4.10: User Rating Behavior based on User Emotion

Finally, we collected users’ feedback on how they overall satisfied with therecommended place list as well, their opinion based on their current emotionas shown in Table 4.4. According to the results, CFEUB and TFEUB listsare highly preferred by users as well matched with their current emotions.

57

Table 4.4: User Satisfaction

Algorithm Overall preference(%) Preference with the emotion (%)

CFN 46.77 -

CFUB 80 -

CFE 60 53.33

CFEUB 87.5 87.5

TFUB 68.75 -

TFE 73.33 80

TFEUB 73.33 86.67

4.7 Conclusion

We derived emotional tags for each selected place of the study based onuser reviews as the effort of adding affective parameters to the dataset. Weemployed Plutchik’s emotion classification for both emotion word detectionderived from reviews and acquired users’ current emotions into the system.Based on the emotion selected by the user, we suggested the recommendedplace list should consist of positive emotion categories: Joy, Anticipation,and Trust.

4.8 Closing Discussion

In this study we examined the effectiveness of using user emotion and behav-ior in context-aware tourist destination recommendation. This was achievedby utilizing Pre-filtering techniques and Tensor factorization. User emotionand behavior are much employed in classical recommendation domains likemusic, movies, and books but the travel destination recommendation is stillin the early stages of development. Since adapting emotion in the con-sumption stage of the recommendation is difficult, in our study we focusedon the effectiveness of using emotion along with user behavior in the pro-posed recommendation. Our system can outperform several state-of-the-artcontext-aware travel recommendation systems. In both Pre-filtering andTensor factorization incorporation of context outperformed well.

58

Chapter 5

Adapting User Personality

and Emotion in

Context-aware Travel

Recommendation

5.1 Introduction

Most of the recommender systems are developed based on user’s previouspreference history captured as ratings given to each item in the system.Context-aware recommendation systems are introduced to provide usersmore relevant information since the recommendation algorithms are not de-rived only with the users’ past preferences as traditional recommendations.Thus, most contextual information are characteristics of an activity suchas location, time and some dynamic features of user profiles such as users’emotional estate and personality traits.

The emotion is one of the popular contextual parameters which isadapted by many RS studies and triggered the effectiveness of emotion inthe recommendation process though the few works have arisen in touristdestination recommendation. The academia in RS has been addressed theincorporation of emotion in recommendation process in terms of the impactof emotion and personality traits in the human decision-making process. Aswell in favor of the user interaction to the RS, the role of emotion has beenidentified in the different stages of the process of RS’s Content Consump-tion as the entry stage, the consumption stage and the exit stage while theconsumption stage is highly influenced by emotion in various domains and

59

led to emerge the area of studies in emotion based RSs to address variousaspects with different types of techniques [96].

Moreover, the users personality can be widely affect users attitude, tastesand behaviors [5]. The most influential model in psychology, namely big fivefactor model which describes the personality according to five traits: open-ness to experience, conscientiousness, extraversion, agreeableness and neu-roticism [28]. In our study we incorporate emotion and personality togetherto evaluate effectiveness in travel destination recommendation. Therefore,our objective is to compare the RSs accuracy and effectiveness based useremotion and personality by using collaborative filtering approach.

5.2 Background and Related Work

Some of the researchers have used users personality in RS because users withdifferent personality traits prefer items with different features which canled to provide valuable information for personalized recommendation [94].Tkalcic et al. [98] proposed an approach for calculating the user similarity toimprove the nearest neighborhood measure in collaborative filtering for RSthat is based on the big five personality model. Hu and Pu [52] address theissue of limiting the success of collaborative filtering in certain applicationdomains is the cold-start problem, a situation where historical data is toosparse, new users have not rated enough items, or both by incorporatinghuman personality into the collaborative filtering framework. Elahi et.al[36] suggested an active learning strategy based on personality, to predictitems that users can rate before they get recommendations.

Emotions are referred as mental states usually caused by an event ofimportance to the subject and in variety of ways emotion has been modeled.The universal model classifies emotion into fixed categories. Ekmans sixbasic emotions: happiness, anger, fear, sadness, disgust and surprise [35]and according to Plutchik, there are eight basic and prototypical emotionswhich are joy,sadness,anger,fear,trust,disgust, surprise, and anticipation.

Adapting the emotion as context parameter into the Recommender Sys-tem is firstly founded by Gonzalez et al (2007)[42]. From then on, manyapplications arisen in the field due to the success of various research work.Music recommendation based on emotion is one of the domain both recom-mendation systems and affective engineering studies are merged [19, 45, 87]due to the diversity and richness of music content and in context-based mu-sic recommendation systems, it often needs multidisciplinary efforts suchas emotion description, emotion detection etc., to the successfulness. Themovie recommendation also enriched with various research efforts by apply-ing emotion in recommendation and few studies emotions was incorporatedin various stages of movie consumption by studies [99, 109, 47].

Yik et al. [102] reported that all five factors of the Five Factor Model

60

influence feelings and emotional behavior. Therefore, in our proposed RSwe try to compare the effectiveness in recommendation by using emotionand personality while incorporating the information of these parameters inuser profiles.

The traveler recommendation systems serve tourists with relevant andpersonalized destination suggestions for helping them making better deci-sions. The state-of-the-art of RSs can be analyzed as web based RSs ormobile based RSs while pointing out the web-based are predominant. Aswell depending on the service, systems offer the recommendation can beclustered as suggestion of a destination and construction of a whole touristpack, recommendation of suitable attractions in one specific destination,design of a detailed multi-day trip schedule, or social capabilities [15].

5.3 Proposed System

The proposed system recommends places with tourist destinations as acontext-aware RS by taking account the users emotion and user person-ality in the recommendation engine. Context acquisition to the system is tobe done in two steps as emotion acquisition and personality acquisition.

5.3.1 Context Acquisition

Emotion Acquisition

Users emotions are extracted in the scale of Happy, Surprise, Disgust,Sad, Afraid, Angry, Confidence and Anticipation based on the Plutchikemotion classification from user’s tweets by employing Sentiment AnalysisTechniques; the Lexicon-based approach which relies on precompiled senti-ment terms.

The emotion lexicon stated by National Research Council Canada (NRC)which is based on Plutchik eight emotions anger, fear, anticipation, trust,surprise, sadness, joy and sadness and two sentiments negative and positivefor English words [70] was used to calculate Term Frequencies.

Figure 5.1: User emotion classification from tweets.

61

The emotion classification process consists of the collection of a userstweets, pre-processing, term frequency (TF) calculation, and daily user emo-tion derivation based on TFs (see Figure 5.1).

The text pre-processing is a much important task in text mining tech-niques and its application as the first step and undergoes several tasks asmentioned below.

� Stop word removal: Removal of words such as the, is, and which

� Tokenization: Each review is broken into its constituent words.

� Conversion to lowercase: Each review text is converted to lowercase.

� Stemming: Reducing tokens into their root words, e.g., words such ashappier, happily, and happiness all map to the root word happy

The main aim behind the pre-processing is to represent each document asa feature vector, so it facilitates to separate the text into individual words[6]. We performed the Term Frequencies (TFs) calculation based on thelexicons of the eight emotion groups as listed in Table 5.1.

Table 5.1: Examples from the NRC Emotion Lexicon.

Emotion Lexicon entries by emotion

Joy absolution, abundance, abundant

Surprise abandonment, abduction, abrupt

Anticipation abundance, accolade, accolade

Trust abacus, abandoned, abandonment

Disgust aberration, abhor, abhorrent

Fear abandon, abandoned, abhor

Anger abandoned, abandonment, abhor

Sadness abandon, abandoned, abduction

According to the maximum frequency value, we chose the emotion cat-egory for a place, as stated in Table 5.2. In the case of the occurrence ofmultiple emotions,we assumed that if an emotion in the negative categoryand an emotion in the positive category appeared with equal frequency, thenegative emotion is stronger, based on the consideration that humans aremore straightforward in reporting negative emotions as compared to posi-tive emotions. In terms of positive emotion categories, in the case of equaloccurrence with a negative emotion, we assumed that Joy, Trust, Antici-pation, and Surprise are stated in descending order by considering the factthat human self-awareness of their emotion at a particular time.

62

Table 5.2: Daily Emotion Derivation for a Tweet.

Emotion Word frequency

Joy 8

Surprise 2

Anticipation 1

Trust 0

Disgust 0

Fear 0

Anger 1

Sadness 0

User emotion Joy

The extracted user emotion is saved to the database and retrieved in therecommendation process to generate recommendations along with the userpersonality.

Personality Acquisition

A User’s personality is detected by analyzing user’s tweets using the IBMWatson Personality Traits service 1.

Figure 5.2: IBM Bluemix to analyze personalities.

The IBM Watson Personality Insights Service derives Insights aboutpersonality characteristics from social media, enterprise data, or other dig-ital communications. The service infers personality characteristics based

1https://www.ibm.com/watson/services/personality-insights/

63

on three models namely Five Big, Needs and Values. The Five Bigpersonality characteristics represent the five factor model characteristics:Agreeableness- person’s tendency to be compassionate and cooperative to-ward others, Conscientiousness- person’s tendency to act in an organized orthoughtful way, Extraversion- person’s tendency to seek stimulation in thecompany of others, Emotional range- referred to as Neuroticism or Naturalreactions , is the extent to which a person’s emotions are sensitive to theperson’s environment and Openness- the extent to which a person is opento experiencing a variety of activities2.

In order to extract a users personality, the users tweets were analyzedby the IBM Watson Personality Insights Service. Based on the results foreach personality trait, the personality category for each user was derived byconsidering the trait for which the highest value was obtained(see Figure 5.2)and was inserted to the user profile, as shown in Figure 5.3. In deriving theusers personality, one week of tweets by the user was collected and analyzed.Most users updated their twitter profiles daily.

Figure 5.3: User profile details for a user.

5.3.2 Recommendation System Implementation

We attempted to incorporate emotion and personality in travel destinationrecommendation based on pre-filtering techniques. In the recommendationprocess, we used user-user collaborative filtering to recommend items. The

2https://console.bluemix.net/docs/services/personality-insights/models

64

similarity computation between items or users is critical in the collabora-tive filtering approach. In user based collaborative filtering, the similarity,Wu,v, is the similarity value between user u and user v, both of whom haverated both items. The similarity calculation can be performed based on dif-ferent types of similarity measures, such as Pearson correlation, Euclideandistance, Tanimoto coefficient, and LogLikelihood similarity. The presentstudy calculated the similarity values using

The Pearson correlation between users u and v is

W u,v =

∑i∈I(r(u,i) − ru)(r(v,i) − rv )√∑

i∈I(r(u,i) − ru)2√∑

i∈I(r(v,i) − rv )2(5.1)

where the i∈ I summations are over the items that both the users u andv have rated and ru is the average rating of the co-rated items of the uth

user.The prediction algorithms guess the rating that a user would provide for

a target item [74]. In order to make a prediction for the active user a anda certain item i, we considered the weighted average of all ratings of thatitem according to the following formula:

P a,i = ra +

∑u∈U (r(u,i) − ru) ·W a,u∑

u∈U |W a,u |(5.2)

where ra and ru are the average ratings for the user a and user u on allother rated items, and wa,u is the weight between the user a and user u.The summations are over all the users u ∈ U who have rated the item i.The top-N recommendation is a set of N number of top-ranked items thatwill be of interest to a certain user and is derived based on the predictiverating values calculated for a user.

5.4 Experimental Set up and Evaluation Protocols

5.4.1 Experimental setup

The Traveler Destinations, place recommendation system was evaluated by18 users and each of them were asked to experiment with RS process andevaluate the recommended lists by rating according to the user preferencefor each place in the list and state the overall preference for the list accordingto users current emotion and overall satisfaction for recommended lists inthe 5 point-Likert Scale.

Figure 5.4 illustrates the architecture of the proposed RS and the con-textual parameters selected for the present study. Emotion and personalitywere used to filter the user ratings as the features of the user profile of

65

Figure 5.4: Proposed recommendation system.

Figure 5.5: Recommended places for a user.

the system, and we optimized the recommendation process by incorporat-ing these contexts to predict more personalized suggestions for the users.The figure 5.5 illustrates a list of recommended places for a user in theimplemented system.

The evaluation was performed to assess the users opinion about the qual-ity of recommendation algorithm CFUFE (collaborative filtering approachwith emotion and personality) to compare baseline approach CFN (collab-orative filtering approach with non-context). The baseline approach doesntincorporate any context in the recommendation algorithm and developedbased on item-item collaborative filtering.

For the baseline approach, the similarity calculation for the recommen-dation process was performed with the loglikelihood ratio, sim(u,n), which isderived based on occurrences related to the users or items for which users oritems overlap in terms of preference(where both compared users have or do

66

not have preferences) and events for which both users or items do not havepreferences. The log-likelihood ratio, sim(u,n), is calculated as follows:

sim(u, n) = −2 lnliklihood for null model

likilihood for alternative model(5.3)

The predictive rating, Pu,i, by user u for item i was done as follows:

P u,i =

∑n∈N sim(u, n)×Ru,n∑n∈N (sim(u, n) + 1)

(5.4)

where sim(u,n) is the similarity between the nth item and user u, andRu,n is the rating by user u of item n for N items based on the Mahout-item-based recommendation algorithm [86]. The similarity calculations reachedthe range from -1.0 to 1.0, and in order to avoid negative values, we added1.0 to the similarity values so that the similarity ranges from 0.0 to 2.0.The list of the top five place recommendations was created based on thehighest predictive rating values in baseline method (CFN) as well. In thefiltering process of recommendation, Disgust, Fear, Anger, and Sadness arecategorized and arranged in three positive groups: Anticipation, Joy, andTrust. In order to avoid negative emotions, Fear and Anger were groupedwith Joy and Trust, respectively.

5.4.2 Evaluation protocols

Evaluation Protocols for the task of recommendation lists, the following areused:

1. We used Precision and Mean Average Precision (MAP) values of thetwo approaches

2. We evaluated the Predictive rating validation of the system by meansof RMSE values for both CFN and CFUFE approaches

3. The overall satisfaction of users towards the recommendation systemanalyzed

In the experiment setup, the testing users had to register to the systemand evaluate the two lists, each with five places while there are postingtweets in their tweeter account.

5.5 Results

In evaluation of recommendation lists, the Precision and Mean Average pre-cision (MAP) values of the two approaches CFN and CFUFE were calculated

67

according to following formulas.

Precision =|Correctly recommended items||Total recommended items|

(5.5)

AveP =

∑nk=1(p(k)× rel(k))

number of relevant items(5.6)

MAP =

∑nk=1AveP (q)

Q(5.7)

where P(k) is the precision at k-th element, rel(k) is 1 if the i-th item ofthe list is relevant and Q is the total no. of lists.Table 5.3 shows the resultsfor the CFUFE and CFN approaches.

Table 5.3: Precision and Mean Average Precision

Algorithm Precision (%) Mean Average Precision (%)

CFN 59.69 61.98

CFUFE 65.49 70.91

The Root Mean Square Error (RMSE) values also are computed for eachapproach and the results are shown below in the figure 5.6. Bi denotes theitems rated by user ai

RMSE =√

(1

|Bi |∑bk∈Bi

(ri(bk )− pi(bk ))2) (5.8)

The CFUFE obtained 0.45 as RMSE score while CFN obtained 2.53 asthe score. The lower results of CFUFE showed better performance in ratingpredictions.

To evaluate user satisfaction for the system, we separately conducteda survey by asking overall satisfaction and their satisfaction towards therecommended lists based on the moods.

Table 5.4: User Satisfaction

Algorithm Overall preference (%) Preference with the emotion (%)

CFN 47 -

CFUFE 63 58

The table 5.4 shows the results of the User Satisfaction for all the userswho participated for the experiment.

68

Figure 5.6: RMSE values for CFUFE and CFN.

5.6 Conclusion

The focus of this chapter is to validate the user emotion in recommendationbased on users personalities. Many recommendation systems have adaptedemotion and personality in recommendation process and very few studiesonly have conducted in traveler destination recommendation domain.This study focuses on a challenging field, such as tourist destinationrecommendation compared to many classical RSs development domainssuch as movies, books, and music and selected the emotion and personalityas contextual parameters.

We incorporated emotion parameter by employing Semantic Analysistechniques on users tweets and derive the daily emotion automatically tothe system based on lexicon based approach and review the effectivenessof incorporating emotion in recommendation process by using pre-filteringtechniques. As well, we review the effectiveness of incorporating person-ality along with emotion in the recommendation process and the pro-posed approach showed better results compared to the baseline approach(MAP=71%).

5.7 Closing Discussion

In this chapter, we show that role of personality as a contextual parameter intourist destination recommendation. Many recommendation systems haveadapted emotion and personality in recommendation process and very few

69

studies only have conducted in traveler destination recommendation domain.To acquire user emotion and personality we used users’ tweets and analyzed.

The current chapter discusses the extension of our study to emotion-based travel recommendation using pre-filtering techniques, where we usethe personality as a context. Furthermore, we rely on tweets in extractingcontexts for the recommendation. We obtained better results in predictiverating calculation as compared to the CFN (RMSE = 0.45), and 63% ofusers overally, preferred the recommendations made by the CFUFE.

70

Chapter 6

Conclusion and Future Work

In this thesis, we have addressed the context-aware Travel Recommendationby means of psychological parameters emotion and personality of users, gen-erated through user reviews and tweets while analyzing the texts. We firstproposed a method of deriving emotion tags by analyzing user reviews forselected destinations from TripAdvisor. Then we used derived emotion tagsincluded dataset in travel recommendation by using pre-filtering techniquesand evaluated the system by comparing a non-context and emotion-based,recommendation. Further, we incorporated the user behavior with emotionand compared the accuracy in the recommendation by using both param-eters and the non-context model by means of pre-filtering and contextualmodeling. Finally, we extended the work of generating user profiles, includ-ing users emotion and personality and evaluated the proposed system notonly in terms of rating prediction and item ranking but also considering theoverall user satisfaction.

6.1 Discussion of Contributions

In chapter 02, we presented a comprehensive survey of the state-of-the-art onemotion and personality based recommendation systems while emphasizingthe recommendation approaches used in the studies. An approach of deriv-ing emotion tags using a text analysis is illustrated in chapter 3 as a result offinding a solution for the key challenge of our study unavailability of the af-fective parameters included context-aware dataset . We could derive emotiontags with the accuracy 63%. We illustrated the accuracy emotion contextapplied travel recommendation with respect to the non-context applied rec-ommendation in travel domain by giving the better accuracy with the Meanaverage precision of 73.8%. In chapter 4, we further analyzed the recom-mendation with user behavior context with emotion for the comparison rea-

71

sons by means of pre-filtering and contextual modeling. In our approachesall the contextual approaches performed with high accuracy compared tothe non-context approach. In every context (single or both contexts) theContextual modeling algorithms outperformed compared to Pre-filtering al-gorithms. As well in each case, when both context applied recommendationalgorithms showed better performances compared to single context whileTensor factorization with emotion and user behavior was the most successful(Mean average precision of 81.2%).The chapter 06 discusses the personalityas contextual parameter with emotion in travel recommendation by usinguser-user collaborative filtering. The recommendation approached reachedto the better performance compared to the baseline approach in MAP valueas 71%.

In implementation of RSs, main focus carries in giving best ’User Expe-rience’ to the user.The User Experience (UX) refers to a persons emotionand attitude about using a particular, system or service. Our study carriedan emotional group analysis on MAP based on Plutchik’s comprehensive listof eight primary emotions arranged as opposing pairs.

We suggested users, who are having negative emotions to shift to thepositive emotion range which is called as ’Context Suggestion’ in the no-tions of RSs where it leads as an emerging research domain. The mainchallenge of the context suggestion, is not having users taste on the contextwe suggest.Thus, we believe MAPE evaluation was able to trace the tasteof users on emotion as context in the RS process.

6.2 Future Work

In deriving emotion tag, we used lexicon-based method since it can enhancewith multiple sources knowledge compared to other approaches in SA andthe accuracy of the emotion tag deriving process can improve by using opin-ion mining, and other SA techniques of the state-of-the-art. Recently, SAtechniques enriched with new trends such as deep learning techniques. TheDeep Convolutional Neural Network and Convolutional Neural Network arefew techniques recently emerged. In our study, we compared recommenda-tion in pre-filtering and contextual modeling approaches with item-item anduser-user collaborative filtering and Tensor factorization. We employed sev-eral parameters in recommendation process as the context namely emotion,user behavior and personality. In our study, we mainly focused on emotionas a contextual parameter, and for the comparison, other contextual param-eters, user behavior and personality also used. Recently, few lexicons ap-peared in SA research works and one such a lexical resource, SentiWordNetwhich is one dictionary of opinionated terms. The senticword is built basedon SentiWordNet lexicon and adapt Hourglass of emotion. In this model,sentiments reorganized around four independent dimensions whose different

72

levels of activation, in fact in this model affective states do not classify intotraditional emotional categories instead into four concomitant but indepen-dent dimensions - Pleasantness, Attention, Sensitivity, and Aptitude. Ourstudy, focused on the primary eight emotions which were derived accordingto Plutchiks emotion classification and chose the NRC lexicon in emotiontag derivation process. The experimental results proved that incorporatingcontextual parameters, enhanced the accuracy of the proposed recommen-dation and out of all the approaches the Tensor Factorization with emotionand user behavior outperformed. The personality was employed to identifyas a contextual parameter in Travel Recommendation and the state-of-the-art of emotion-based and personality-based recommendation systems theimplication of user personality on emotion in the recommendation processhas been discussed based on the theories of psychological explanations onemotion. Therefore, we recommend tracing the personality implication onemotion for the average precision of a user in the process of interaction withRS for further studies.

Table 6.1: Personality implication on emotion

User ID Emotion Personality Av.Precision (%)

30 Joy Openness 0.89

37 Trust Emotional range 1.00

9 Anticipation Conscientiousness 0.58

29 Disgust Conscientiousness 0.60

28 Joy Conscientiousness 0.48

Further, any quantitative result that would prove, any of the personalitytraits implication on emotion or emotions would open a new direction forcontext-aware recommendation studies that suggest novel Similarity Mea-sure for an emotion-based recommendation with pre-filtering techniques.For further studying, employing these contextual parameters in contextualmodeling approaches like dependent context-aware approach categories (e.g.,deviation-based or similarity-based) will expose a more appropriate methodon travel recommendation.

73

Publications

The work presented here in this thesis has appeared in two journals andtwo conference proceedings. We list these publications as below.

Journal Articles

1. Ishanka, U. A., & Yukawa, T. (2017). The Prefiltering Techniquesin Emotion Based Place Recommendation Derived by User Reviews.Applied Computational Intelligence and Soft Computing, 2017.

2. Ishanka, U. A., & Yukawa, T. (2018). An Analysis of Emotion and userBehavior in Context-Aware Travel Recommendation Systems usingPre-Filtering and Tensor Factorization Techniques. Global Journal ofComputer Science and Technology: D Neural & Artificial Intelligence,2018.

Conference Proceedings

1. Ishanka UAP, Takashi Yukawa. (2017). A Context-aware Place Rec-ommendation System based on Emotion and User Behavior as Contex-tual Parameters. The 18th International Symposium on Advanced In-telligent Systems (ISIS 2017): Springer International Publishing, 2017.

2. UA Piumi Ishanka , Takashi Yukawa. (2018). User Emotion and Per-sonality in Context-aware Travel Destination Recommendation. 5th

International Conference on Advance Informatics: Concepts, Theoryand Applications(ICAICTA 2018): IEEE, 2018.(not yet published)

74

Bibliography

[1] Emotions and Personality in Personalized Services: Models, Evalua-tion and Applications, chapter Emotions in Context-Aware Recom-mender Systems. Springer International Publishing, 2016.

[2] Gediminas Adomavicius, Ramesh Sankaranarayanan, Shahana Sen,and Alexander Tuzhilin. Incorporating contextual information in rec-ommender systems using a multidimensional approach. ACM Trans-actions on Information Systems (TOIS), 23(1):103–145, 2005.

[3] Gediminas Adomavicius and Alexander Tuzhilin. Toward the nextgeneration of recommender systems: A survey of the state-of-the-artand possible extensions. IEEE transactions on knowledge and dataengineering, 17(6):734–749, 2005.

[4] Gediminas Adomavicius and Alexander Tuzhilin. Context-aware rec-ommender systems. In Recommender systems handbook, pages 191–226. Springer, 2015.

[5] Icek Ajzen. Attitudes, personality, and behavior. McGraw-Hill Educa-tion (UK), 2005.

[6] Saima Aman and Stan Szpakowicz. Identifying expressions of emotionin text. In International Conference on Text, Speech and Dialogue,pages 196–205. Springer, 2007.

[7] Oscar Araque, Ignacio Corcuera-Platas, J Fernando Sanchez-Rada,and Carlos A Iglesias. Enhancing deep learning sentiment analysiswith ensemble techniques in social applications. Expert Systems withApplications, 77:236–246, 2017.

[8] Daniar Asanov et al. Algorithms and methods in recommender sys-tems. Berlin Institute of Technology, Berlin, Germany, 2011.

[9] Saikat Bagchi. Performance and quality assessment of similarity mea-sures in collaborative filtering using mahout. Procedia Computer Sci-ence, 50:229–234, 2015.

75

[10] Linas Baltrunas and Xavier Amatriain. Towards time-dependant rec-ommendation based on implicit feedback. In Workshop on context-aware recommender systems (CARS09), 2009.

[11] Manuel J Barranco, Jose M Noguera, Jorge Castro, and Luis Martınez.A context-aware mobile recommender system based on locationand trajectory. In Management intelligent systems, pages 153–162.Springer, 2012.

[12] John R. L Bernard. The Macquarie thesaurus : the book of words. DeeWhy, NSW : Macquarie Library Pty. Ltd, rev. ed edition, 1986. 1999reprint has subtitle: Australia’s national thesaurus.

[13] C Bishop. Bishop cm: Pattern recognition and machine learning.springer. Journal of Electronic Imaging, 16(4):140–155, 2006.

[14] Matthias Bohmer, Gernot Bauer, and Antonio Kruger. Exploring thedesign space of context-aware recommender systems that suggest mo-bile applications. In 2nd Workshop on Context-Aware RecommenderSystems, volume 5, 2010.

[15] Joan Borras, Antonio Moreno, and Aida Valls. Intelligent tourismrecommender systems: A survey. Expert Systems with Applications,41(16):7370–7389, 2014.

[16] Matthias Braunhofer, Victor Codina, and Francesco Ricci. Switchinghybrid for cold-starting context-aware recommender systems. In Pro-ceedings of the 8th ACM Conference on Recommender systems, pages349–352. ACM, 2014.

[17] Matthias Braunhofer, Mehdi Elahi, Mouzhi Ge, and Francesco Ricci.Context dependent preference acquisition with personality-based ac-tive learning in mobile recommender systems. In International Con-ference on Learning and Collaboration Technologies, pages 105–116.Springer, 2014.

[18] John S Breese, David Heckerman, and Carl Kadie. Empirical analysisof predictive algorithms for collaborative filtering. In Proceedings of theFourteenth conference on Uncer-tainty in artificial intelligence, pages43–52. Morgan Kaufmann Publishers Inc., 1998.

[19] Rui Cai, Chao Zhang, Chong Wang, Lei Zhang, and Wei-Ying Ma.Musicsense: contextual music recommendation using emotional allo-cation modeling. In Proceedings of the 15th ACM international con-ference on Multimedia, pages 553–556. ACM, 2007.

[20] Erik Cambria. Affective computing and sentiment analysis. IEEEIntelligent Systems, 31(2):102–107, 2016.

76

[21] Erik Cambria, Soujanya Poria, Rajiv Bajpai, and Bjorn Schuller. Sen-ticnet 4: A semantic resource for sentiment analysis based on concep-tual primitives. In Proceedings of COLING 2016, the 26th Interna-tional Conference on Computational Linguistics: Technical Papers,pages 2666–2677, 2016.

[22] Pedro G Campos, Ignacio Fernandez-Tobıas, Ivan Cantador, and Fer-nando Dıez. Context-aware movie recommendations: an empiricalcomparison of pre-filtering, post-filtering and contextual modeling ap-proaches. In International Conference on Electronic Commerce andWeb Technologies, pages 137–149. Springer, 2013.

[23] Lineu Castello. Rethinking the meaning of place: conceiving place inarchitecture-urbanism. Routledge, 2016.

[24] Luis Castillo, Eva Armengol, Eva Onaindıa, Laura Sebastia, JesusGonzalez-Boticario, Antonio Rodrıguez, Susana Fernandez, Juan DArias, and Daniel Borrajo. Samap: An user-oriented adaptive sys-tem for planning tourist visits. Expert Systems with Applications,34(2):1318–1332, 2008.

[25] Tomas Chamorro-Premuzic and Adrian Furnham. Personality and mu-sic: Can traits explain how people use music in everyday life? BritishJournal of Psychology, 98(2):175–185, 2007.

[26] Gokul Chittaranjan, Jan Blom, and Daniel Gatica-Perez. Mininglarge-scale smartphone data for personality studies. Personal andUbiquitous Computing, 17(3):433–450, 2013.

[27] Pierre Comon. Tensors: a brief introduction. IEEE Signal ProcessingMagazine, 31(3):44–53, 2014.

[28] JR Costa and T PAUL. of personality theories: Theoretical contextsfor the five-factor model. The five-factor model of personality: Theo-retical perspectives, pages 51–87, 1996.

[29] Toon De Pessemier, Jeroen Dhondt, Kris Vanhecke, and Luc Martens.Travelwithfriends: a hybrid group recommender system for traveldestinations. In Workshop on Tourism Recommender Systems(TouRS15), in conjunction with the 9th ACM Conference on Recom-mender Systems (RecSys 2015), pages 51–60, 2015.

[30] Guy Debord. Introduction to a critique of urban geography, les levresnues# 6, 1955.

[31] Anind K Dey. Understanding and using context. Personal and ubiq-uitous computing, 5(1):4–7, 2001.

77

[32] Marcos A Domingues, Alıpio Mario Jorge, and Carlos Soares. Us-ing contextual information as virtual items on top-n recom-mendersystems. arXiv preprint arXiv:1111.2948, 2011.

[33] Richard O Duda, Peter E Hart, and David G Stork. Pattern classifi-cation. John Wiley & Sons, 2012.

[34] Ted Dunning. Accurate methods for the statistics of surprise andcoincidence. Computational linguistics, 19(1):61–74, 1993.

[35] Paul Ekman. Basic emotions in dalgleish t. e power t.(eds.), the hand-book of cognition and emotion, 1999.

[36] Mehdi Elahi, Matthias Braunhofer, Francesco Ricci, and MarkoTkalcic. Personality-based active learning for collaborative filteringrecommender systems. In Congress of the Italian Association for Ar-tificial Intelligence, pages 360–371. Springer, 2013.

[37] Richard M Felder, Linda K Silverman, et al. Learning and teachingstyles in engineering education. Engineering education, 78(7):674–681,1988.

[38] Nir Friedman, Dan Geiger, and Moises Goldszmidt. Bayesian networkclassifiers. Machine learning, 29(2-3):131–163, 1997.

[39] Damianos Gavalas and Michael Kenteris. A web-based pervasive rec-ommendation system for mobile tourist guides. Personal and Ubiqui-tous Computing, 15(7):759–770, 2011.

[40] Ken Goldberg, Theresa Roeder, Dhruv Gupta, and Chris Perkins.Eigentaste: A constant time collaborative filtering algorithm. infor-mation retrieval, 4(2):133–151, 2001.

[41] Songjie Gong. A collaborative filtering recommendation algorithmbased on user clustering and item clustering. JSW, 5(7):745–752, 2010.

[42] Gustavo Gonzalez, Josep Lluıs De La Rosa, Miquel Montaner, andSonia Delfin. Embedding emotional context in recommender systems.In Data Engineering Workshop, 2007 IEEE 23rd International Con-ference on, pages 845–852. IEEE, 2007.

[43] Lars Grasedyck, Daniel Kressner, and Christine Tobler. A litera-ture survey of low-rank tensor approximation techniques. GAMM-Mitteilungen, 36(1):53–78, 2013.

[44] Zahra Yusefi Hafshejani, Marjan Kaedi, and Afsaneh Fatemi. Improv-ing sparsity and new user problems in collaborative filtering by clus-tering the personality factors. Electronic Commerce Research, pages1–24.

78

[45] Byeong-jun Han, Seungmin Rho, Sanghoon Jun, and Eenjun Hwang.Music emotion classification and context-based music recommenda-tion. Multimedia Tools and Applications, 47(3):433–460, 2010.

[46] Jonathan L Herlocker, Joseph A Konstan, Loren G Ter-veen, andJohn T Riedl. Evaluating collaborative filtering recommender systems.ACM Transactions on Information Systems (TOIS), 22(1):5–53, 2004.

[47] Ai Thanh Ho, Ilusca LL Menezes, and Yousra Tagmouti. E-mrs:Emotion-based movie recommender system. In Proceedings of IADISe-Commerce Conference. USA: University of Washington Both-ell,pages 1–8, 2006.

[48] John L Holland. Making vocational choices: A theory of vocationalpersonalities and work environments. Psychological Assessment Re-sources, 1997.

[49] Meng-Yen Hsieh, Wen-Kuang Chou, and Kuan-Ching Li. Buildinga mobile movie recommendation service by user rating and app us-age with linked data on hadoop. Multimedia Tools and Applications,76(3):3383–3401, 2017.

[50] Rong Hu and Pearl Pu. A study on user perception of personality-based recommender systems. In International Conference on UserModeling, Adaptation, and Personalization, pages 291–302. Springer,2010.

[51] Rong Hu and Pearl Pu. Using personality information in collaborativefiltering for new users. Recommender Systems and the Social Web, 17,2010.

[52] Rong Hu and Pearl Pu. Enhancing collaborative filtering systems withpersonality information. In Proceedings of the fifth ACM conferenceon Recommender systems, pages 197–204. ACM, 2011.

[53] Ya-Han Hu, Pei-Ju Lee, Kuanchin Chen, J Michael Tarn, and Duyen-Vi Dang. Hotel recommendation system based on review and contextinformation: a collaborative filtering appro. In PACIS, page 221, 2016.

[54] Yuxia Huang and Ling Bian. A bayesian network and analytic hier-archy process based personalized recommendations for tourist attrac-tions over the internet. Expert Systems with Applications, 36(1):933–943, 2009.

[55] FO Isinkaye, YO Folajimi, and BA Ojokoh. Recommendation systems:Principles, methods and evaluation. Egyptian Informatics Journal,16(3):261–273, 2015.

79

[56] John A Johnson. Descriptions used in ipip-neo narrative report. Re-trieved December, 2009.

[57] Marius Kaminskas and Francesco Ricci. Emotion-based matching ofmusic to places. In Emotions and Personality in Personalized Services,pages 287–310. Springer, 2016.

[58] David Keirsey. Please understand me 2. Prometheus Nemesis BookCompany, 1998.

[59] Tamara G Kolda and Brett W Bader. Tensor decompositions andapplications. SIAM review, 51(3):455–500, 2009.

[60] Michal Kompan and Maria Bielikova. Social structure and personalityenhanced group recommendation. In UMAP Workshops, 2014.

[61] Joseph A Konstan, Bradley N Miller, David Maltz, Jonathan L Her-locker, Lee R Gordon, and John Riedl. Grouplens: applying collabora-tive filtering to usenet news. Communications of the ACM, 40(3):77–87, 1997.

[62] F Lemos, R Carmo, Windson Viana, and R Andrade. Towards acontext-aware photo recommender system. In Context-aware recom-mender system workshops, 2012.

[63] Sabrina Lombardi, Sarabjot Singh Anand, and M Gorgoglione. Con-text and customer behaviour in recommendation. 2009.

[64] Marc Moreno Lopez and Jugal Kalita. Deep learning applied to nlp.arXiv preprint arXiv:1703.03091, 2017.

[65] Fabiana G Marinho, Rossana MC Andrade, Claudia Werner, WindsonViana, Marcio EF Maia, Lincoln S Rocha, Eldanae Teixeira, Joao BFerreira Filho, Valeria LL Dantas, Fabrıcio Lima, et al. Mobiline: Anested software product line for the domain of mobile and context-aware applications. Science of Computer Programming, 78(12):2381–2398, 2013.

[66] Sandra Matz, Yin Wah Fiona Chan, and Michal Kosinski. Modelsof personality. In Emotions and Personality in Personalized Services,pages 35–54. Springer, 2016.

[67] Robert R McCrae and Oliver P John. An introduction to the five-factor model and its applications. Journal of personality, 60(2):175–215, 1992.

[68] Walaa Medhat, Ahmed Hassan, and Hoda Korashy. Sentiment anal-ysis algorithms and applications: A survey. Ain Shams EngineeringJournal, 5(4):1093–1113, 2014.

80

[69] Soha A Mohamed, Taysir Hassan A Soliman, and Adel A Sewisy. Acontext-aware recommender system for personalized places in mobileapplications. Int. J. Adv. Comput. Sci. Appl, 7(3):442–448, 2016.

[70] Saif M Mohammad and Peter D Turney. Crowdsourcing a word–emotion association lexicon. Computational Intelligence, 29(3):436–465, 2013.

[71] Mohamed Mostafa, Tom Crick, Ana C Calderon, and Giles Oatley.Incorporating emotion and personality-based analysis in user-centeredmodelling. In International Conference on Innovative Techniques andApplications of Artificial Intelligence, pages 383–389. Springer, 2016.

[72] Jose M Noguera, Manuel J Barranco, Rafael J Segura, and LuisMartıNez. A mobile 3d-gis hybrid recommender system for tourism.Information Sciences, 215:37–52, 2012.

[73] Keith Oatley, Dacher Keltner, and Jennifer M Jenkins. Understandingemotions. Blackwell publishing, 2006.

[74] Manos Papagelis and Dimitris Plexousakis. Qualitative analysis ofuser-based and item-based prediction algorithms for recommendationagents. Engineering Applications of Artificial Intelligence, 18(7):781–789, 2005.

[75] Doo-Soon Park. Improved movie recommendation system based-onpersonal propensity and collaborative filtering. KIPS transactions oncomputer and communication systems, 2(11):475–482, 2013.

[76] Biljana Petrevska and Saso Koceski. Tourism recommendation system:empirical investigation. Revista de turism-studii si cercetari in turism,(14):11–18, 2012.

[77] Robert Plutchik and Henry Kellerman. Emotion, Theory, Research,and Experience: Theory, Research and Experience. Academic press,1980.

[78] Soujanya Poria, Erik Cambria, and Alexander Gelbukh. Deep convolu-tional neural network textual features and multiple kernel learning forutterance-level multimodal sentiment analysis. In Proceedings of the2015 conference on empirical methods in natural language processing,pages 2539–2544, 2015.

[79] Lara Quijano-Sanchez, Juan A Recio-Garcia, and Belen Diaz-Agudo.Personality and social trust in group recommendations. In Tools withArtificial Intelligence (ICTAI), 2010 22nd IEEE International Con-ference on, volume 2, pages 121–126. IEEE, 2010.

81

[80] Peter J Rentfrow and Samuel D Gosling. The do re mi’s of everydaylife: The structure and personality correlates of music preferences.Journal of personality and social psychology, 84(6):1236, 2003.

[81] Paul Resnick and Hal R Varian. Recommender systems. Communica-tions of the ACM, 40(3):56–58, 1997.

[82] Francesco Ricci, Lior Rokach, and Bracha Shapira. Introduction torecommender systems handbook. In Recommender systems handbook,pages 1–35. Springer, 2011.

[83] Alexandra Roshchina, John Cardiff, and Paolo Rosso. A comparativeevaluation of personality estimation algorithms for the twin recom-mender system. In Proceedings of the 3rd international workshop onSearch and mining user-generated contents, pages 11–18. ACM, 2011.

[84] Maede Kiani Sarkaleh, Mehregan Mahdavi, and Mahsa Baniardalan.Designing a tourism recommender system based on location, mobiledevice and user features in museum. International Journal of Manag-ing Information Technology, 4(2):13, 2012.

[85] Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl.Item-based collaborative filtering recommendation algorithms. In Pro-ceedings of the 10th international conference on World Wide Web,pages 285–295. ACM, 2001.

[86] Sebastian Schelter and Sean Owen. Collaborative filtering with apachemahout. Proc. of ACM RecSys Challenge, 2012.

[87] Man-Kwan Shan, Fang-Fei Kuo, Meng-Fen Chiang, and Suh-Yin Lee.Emotion-based music recommendation by affinity discovery from filmmusic. Expert systems with applications, 36(4):7666–7674, 2009.

[88] Guy Shani and Asela Gunawardana. Evaluating recommendation sys-tems. In Recommender systems handbook, pages 257–297. Springer,2011.

[89] Bart Stewart. Personality and play styles: A unified model. Gamasu-tra, September, 1, 2011.

[90] Xiaoyuan Su and Taghi M Khoshgoftaar. A survey of collaborativefiltering techniques. Advances in artificial intelligence, 2009:4, 2009.

[91] Huifeng Tang, Songbo Tan, and Xueqi Cheng. A survey on sentimentdetection of reviews. Expert Systems with Applications, 36(7):10760–10773, 2009.

[92] Kenneth W Thomas. Conflict and conflict management: Reflectionsand update. Journal of organizational behavior, 13(3):265–274, 1992.

82

[93] Nava Tintarev, Matt Dennis, and Judith Masthoff. Adapting rec-ommendation diversity to openness to experience: A study of humanbehaviour. In International Conference on User Modeling, Adaptation,and Personalization, pages 190–202. Springer, 2013.

[94] Marko Tkalcic and Li Chen. Personality and recommender systems.In Recommender systems handbook, pages 715–739. Springer, 2015.

[95] Marko Tkalcic, Berardina De Carolis, Marco de Gemmis, Ante Odic,and Andrej Kosir. Empire 2013: Emotions and personality in person-alized services.

[96] Marko Tkalcic, Andrej Kosir, and Jurij Tasic. Affective recommendersystems: the role of emotions in recommender systems. In Proc. TheRecSys 2011 Workshop on Human Decision Making in RecommenderSystems, pages 9–13. Citeseer, 2011.

[97] Marko Tkalcic, Andrej Kosir, and Jurij Tasic. The ldos-peraff-1 corpusof facial-expression video clips with affective, personality and user-interaction metadata. Journal on Multimodal User Interfaces, 7(1-2):143–155, 2013.

[98] Marko Tkalcic, Matevz Kunaver, Jurij Tasic, and Andrej Kosir. Per-sonality based user similarity measure for a collaborative recommendersystem. In Proceedings of the 5th Workshop on Emotion in Human-Computer Interaction-Real world challenges, pages 30–37, 2009.

[99] Karzan Wakil, Rebwar Bakhtyar, Karwan Ali, and Kozhin Alaadin.Improving web movie recommender system based on emotions. In-ternational Journal of Advanced Computer Science and Applications,6(2), 2015.

[100] Jianwang Wang. A collaborative filtering systems based on person-ality information. In 2015 International Industrial Informatics andComputer Engineering Conference, Shaanxi, China, January, pages10–11, 2015.

[101] Wen Wu, Li Chen, and Liang He. Using personality to adjust diversityin recommender systems. In Proceedings of the 24th ACM conferenceon hypertext and social media, pages 225–229. ACM, 2013.

[102] Michelle SM Yik, James A Russell, Chang-kyu Ahn, Jose Miguel Fer-nandez Dols, and Naoto Suzuki. Relating the five-factor model ofpersonality to a circumplex model of affect. In The five-factor modelof personality across cultures, pages 79–104. Springer, 2002.

83

[103] Soe Tsyr Yuan and Chun-Ya Yang. Service recommender system basedon emotional features and social interactions. Kybernetes, 46(2):236–255, 2017.

[104] Yong Zheng. Deviation-based and similarity-based contextual slimrecommendation algorithms. In Proceedings of the 8th ACM Confer-ence on Recommender Systems, RecSys ’14, pages 437–440, New York,NY, USA, 2014. ACM.

[105] Yong Zheng. Adapt to emotional reactions in context-aware person-alization. In EMPIRE@ RecSys, pages 1–8, 2016.

[106] Yong Zheng, Robin Burke, and Bamshad Mobasher. Differential con-text relaxation for context-aware travel recommendation. In Inter-national Conference on Electronic Commerce and Web Technologies,pages 88–99. Springer, 2012.

[107] Yong Zheng, Bamshad Mobasher, and Robin Burke. Carskit: A java-based context-aware recommendation engine. In Data Mining Work-shop (ICDMW), 2015 IEEE International Conference on, pages 1668–1671. IEEE, 2015.

[108] Yong Zheng, Bamshad Mobasher, and Robin Burke. Similarity-basedcontext-aware recommendation. In Proceedings, Part I, of the 16thInternational Conference on Web Information Systems Engineering— WISE 2015 - Volume 9418, pages 431–447, New York, NY, USA,2015. Springer-Verlag New York, Inc.

[109] Yong Zheng, Bamshad Mobasher, and Robin D Burke. The roleof emotions in context-aware recommendation. Decisions@ RecSys,2013:21–28, 2013.

84