HAL Id: hal-01285971https://hal.archives-ouvertes.fr/hal-01285971

Submitted on 10 Mar 2016

HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.

Reality mining: digging the impact of friendship andlocation on crowd behavior

Yuanfang Chen, Antonio M. Ortiz, Noel Crespi, Lei Shu, Lin Lv

To cite this version:Yuanfang Chen, Antonio M. Ortiz, Noel Crespi, Lei Shu, Lin Lv. Reality mining: digging the impactof friendship and location on crowd behavior. MOBIQUITOUS 2013 : 10th International Conferenceon Mobile and Ubiquitous Systems: Computing, Networking and Services, Dec 2013, Tokyo, Japan.pp.142 - 154, 10.1007/978-3-319-11569-6_12. hal-01285971

Reality Mining: Digging the Impact of Friendship andLocation on Crowd Behavior

Yuanfang Chen1, Antonio M. Ortiz1, Noel Crespi1, Lei Shu2, and Lin Lv3

1 Institut Mines-Telecom, Telecom SudParis, France2 Guangdong University of Petrochemical Technology, China3 School of Software, Dalian University of Technology, China

yuanfang.chen,antonio.ortiz_torres,noel.crespi@telecom-sudparis.eu,lei.

shu@lab.gdupt.edu.cn,lvlin1023@gmail.com

Abstract. Crowd behavior is a natural instinct of human, which directly impactshow we form opinions and make decisions. It is a subject that deserves to be stud-ied since it is common that people change their behavior when being in a group.In pervasive computing research, plenty of work has been directed towards dis-covering human movement patterns based on wireless networks, mainly focusingon movements of individuals. It is surprising that social interactions among in-dividuals in a crowd is largely neglected. Mobile phones offer on-body trackingand they are already deployed on a large scale, allowing the characterization ofuser behavior through large amounts of wireless information collected by mo-bile phones. In this paper, we observe and analyze the impact of friendship andlocation attributes on crowd behavior, using location-based wireless mobility in-formation. This is a cornerstone for predicting crowd behavior, which can be usedin a large number of applications such as traffic management, crowd safety, andinfrastructure deployment.

Key words: Crowd behavior, Mobile phones, Wearable computing, Complex so-cial networks

1 Introduction

With the increasing size and frequency of mass events, such as traffic congestion ona highway, swarming at a tourist attraction, and clogging at weekend shopping sale,the study of crowd dynamics has become an important research area [28, 7]. Howev-er, even successful modeling approaches such as those inspired by Newtonian forcemodels are still not fully consistent with empirical observations and are sometimes d-ifficult to be adapted for crowd prediction. With the prevalence of smart devices (suchas Smart Cellphone, Tablet PC, etc.), on-body sensing, computing and communicationhave become widespread [16] (carry-on smart devices can be called “social sensors”).These developments have made it possible to obtain real-time and comprehensive em-pirical data required by crowd behavior analyses [24]. This “reality mining” [18] isdeemed adequate to provide objective measurements of human interactions, and it canbe called “honest signals” [25]. These options open up new pathways in ComputationalSocial Science [17], where large amounts of information obtained from wireless mo-

2 Yuanfang Chen et al.

bile devices can lead to new perspectives for the analysis of crowd behavior and socialdynamics.

In order to develop reliable prediction models for traffic management, urban infras-tructure deployment, or crowd safety, it is necessary to understand what laws determinethe formation of a crowd. While a lot of studies know the “physics” of crowd behavior,it is surprising that social interactions among individuals in a crowd has been large-ly neglected. Indeed, the great majority of existing studies investigate a crowd as acollection of isolated individuals, and each individual has its own motion speed anddirection [20, 1]. However, it turns out that in practice, the majority of individuals donot take actions alone, but in groups with social relationships [26, 10].

In this paper, we first focus on recognizing human crowd behavior by analyzing thedata measured by internet-accessible mobile phones from a location-aware online socialnetwork. By crowd behavior recognition, we understand that the movement of a largenumber of individuals has a pattern and can be attributed, depending on relevant param-eters such as the friendship between individuals and check-in locations (with time) ofthese individuals. Based on the recognition, we can realize that it is possible to predictthe formation of a crowd from wireless information related to individuals. For instance,the formation of downtown pedestrian flows is related to the interactions between indi-viduals, and these flows can be distinguished by collecting the wireless information ofpedestrians.

Second, we investigate how “friendship” and “location” impact human crowd be-havior. Nathan Eagle et al. have obtained an important result: “Data collected frommobile phones has the potential to provide insight into the relational dynamics of indi-viduals. Furthermore, it is possible to accurately infer 95% of friendship relations onlyconsidering the observational wireless data” [9]. This implies that there is a relationshipbetween the friendship and the behavioral patterns of humans. Based on this result, weobserve the patterns of friendship in different crowds.

The contributions of this paper are listed as follows.

– We design a crowd recognition model. One of the main challenges in crowd behaviorrecognition is to infer the most likely crowd behavior using the data collected from aset of persons. We use check-in time and location (Time and Location id. We converteach latitude/longitude coordinate of the earth into a unique Location id) to quantifythe track of each individual. Then, a clustering algorithm1 is used to find the likelycrowds.

– We investigate how friendship influences crowd behavior. In order to measure thisinfluence, we use the friendship degree of each user and the probability distributionof various degrees.

– We investigate how individuals’ locations influence crowd behavior. For measuringthe influence, we investigate these relationships for different clusters: (i) users andtheir locations; (ii) check-in time and users’ locations.

1 An Expectation-Maximization (EM) clustering algorithm is used in this paper. The EM assignsa probability distribution for each track record (instance), which indicates the probability ofeach instance belonging to each of the clusters. The EM can automatically decide how manyclusters to create.

Crowd Behavior 3

The paper is structured as follows. Section 2 introduces the related work. Section 3presents the data used in our study and statistic analysis. Section 4 shows how to rec-ognize human crowd behavior from datasets. Section 5 reveals the impact of attributes(check-in time, check-in location and friendship) on crowd behavior. Finally, some con-clusions are given in Section 6.

2 Related Work

Recently, a number of scientific communities, from computer science to physics, havebeen working in human dynamics. Pedestrian movement patterns have been studiedusing wireless-based personal location data [23]. In physics, for human crowd behav-ior analysis, many approaches have been proposed inspired by using fluid dynamic-s [13], swarms [3] and cellular automata [2]. Car-based human movement patterns havealso been studied by utilizing the data from GPS-equipped vehicles [27]. Thereinto,the approaches which are based on wireless mobility information are more reliable,objective and environment-independent compared with model-based approaches, e.g.,habit-based model [15]. Model-based approaches are environment-sensitive. Moreover,because the factors of environment influence each other, a simple model is not suffi-cient to reflect the interactions between these factors. Furthermore, the performance ofa model is related to the experience of modeler.

As an important aspect of human dynamics, “crowd dynamics” is worthy to bedeeply analyzed, since it helps to extract some useful conclusions about how human-s behave when they are in large groups. Further, predicting the formation of a crowdis helpful in some emergency situations, e.g., evacuation route control; and even theprediction is also beneficial for studying and improving the performance of public in-frastructures, e.g., network usage during a mass event. However, the prediction of crowdevents is still a challenge, even if a lot of new technologies can be used, e.g., GPS-basedhuman trace tracking technology. Moreover, GSM, bluetooth or WiFi localization tech-nologies have been explored to be used to collect sufficient data for analyzing crowdbehavior [6].

The prediction of human behavior is the main topic of a number of publication-s. The method proposed in [14] estimates an object’s future locations, by consideringthe patterns of recent movements. They present the concept of a Trajectory pattern (T-pattern), a special association rule with a timestamp that is able to define a sequence oflocations with a certain probability. They also propose the Trajectory Pattern Tree (TP-T), a data access method that indexes trajectory patterns to efficiently answer predictivequeries, and finally, they detail a Hybrid Prediction Algorithm (HPA) that provides ac-curate prediction for both near and distant time queries. T-patterns are also used in [19],where WhereNext is presented, which is a technique to predict the next location of amoving object. It uses an evaluation function that efficiently creates TPTs by consid-ering the previous movements of all moving objects in a certain area. The model pre-sented in [21] is based on behavioral heuristics that predict individual trajectories andcollective motion patterns such as the spontaneous formation of unidirectional lanes orstop-and-go waves. These heuristics consider visual information to describe the motionof pedestrians.

4 Yuanfang Chen et al.

Most of the previous research in crowd dynamics has ignored the internal connec-tion between social relationships and crowd formation. While discrete observations ofan individual’s idiosyncratic behavior seem to be merely random, and most studies ofcrowd behavior consider only interactions among isolated individuals, the results pre-sented in [22] show that up to 70% of people in a crowd are actually moving in group-s such as friends, couples, or families, concluding that the social relationships affectcrowd forming. In addition, group sizes are commonly distributed according to a pois-son distribution [12]. Thus, social ties between individuals impact the forming of crowd.In this work, we consider friendship and location relationships between individuals asan important parameter for crowd recognition.

3 Data Description

The dataset used for analyzing crowd behavior consists of anonymous check-in datafrom mobile devices collected by a location-based social networking service providerwhere users share their locations by check-in, and the friendship between these mobileusers is collected using their public API [5]. This aggregated and anonymous mobiledevice information is used to correlate, model, evaluate and analyze the relationshipsbetween the check-in time, locations, friendship and crowd behavior of users in 772, 966distinct places. The dataset consists of 58, 228 nodes (users) and 214, 078 friend edges(friendship is directed between any two nodes).

Check-in behavior of users. Based on users’ check-in behavior, we can obtainusers’ locations, and our recognition and observation are check-in-location-aware.Moreover, through check-in frequency, some special places can be inferred. For ex-ample, the home location of a user can be defined as the location which has maximumaverage number of check-ins for a period of time. Manual inspection shows that thisinfers home locations with 85% accuracy [5]. We can deduce that there is a relation-ship between users’ social relations (e.g., the kinship of users at home) and check-inlocations with check-in frequency.

Friendship and mobility. Our study aims at understanding how the location ofuser A’s friend B affects the movement of A. Intuitively we are more likely to moveto a place in which we have friends (crowd behavior is a natural instinct of human).To quantify this effect we proceed as follows. User A “visits” the location of friendB, if A checks in within radius r of B’s location, and we aim at computing p(d), whichmeasures the probability that A “visits” a friend (near the friend) within the range, r = d(we set r = 10m). In the dataset, using the “location id”, we can draw the “ranges”corresponding to different friends and obtain relevant probabilities to which range Abelongs. Moreover, In Fig. 1, we show, for each user, the number of records and towhich cluster belongs based on January, 2010 dataset (Fig. 1(a)) and February, 2010dataset (Fig. 1(b)).

The original dataset provided by the service provider can not be directly used tomine the basic laws that govern human crowd behavior, or even the impact of theselaws, so we apply a process to perform an association of the mobility check-in datawith the friendship information. The process involves two steps:

Crowd Behavior 5

0 1 2 3 4 5 6x 104

0

50

100

150

200

250

300

UserIDN

umbe

r of r

ecor

ds

(a) Result of January, 2010dataset.

0 1 2 3 4 5 6x 104

0

50

100

150

200

250

300

350

400

450

500

UserID

Num

ber o

f rec

ords

(b) Result of February, 2010dataset.

Fig. 1. Number of records and which cluster belongs for each user in the January and February,2010 datasets (different colours represent different clusters).

1. Perform a spatial-temporal analysis of the records to detect which users form acrowd (e.g., places in which some people have stopped for a sufficiently long time);

2. Infer in every crowd which users are friends and what kind of friendship they have(1, 2, 3, ...,N-hop friends, e.g., “1-hop” is for direct friends).

In order to infer “where and when crowds are forming?” from a large number ofrecords, we first characterize the individual activity by mathematical modeling. Eachlocation measurement mi, collected for every mobile device, is characterized by a po-sition pi expressed in latitude and longitude, and a timestamp ti. We also measure theinterval time between different check-in activities [11]. The average interval time mea-sured for the whole population is 30 minutes. So within the interval we can detect agathering of humans from the dataset. Moreover, this time interval is comparable to theaverage length of real social events.

To confirm the friendship in a crowd, we merge the friendship dataset with thecheck-in record dataset, e.g., for a crowd, if there is a record of any two users in thefriendship dataset, the two users are 1-hop friends, and then we add this information intothe check-in record dataset as a value of a new attribute column (friendship). Multihopfriendship will be also recognized, and we record the identity number of minimum-hop-count friend for every user as the value of the friendship attribute column.

4 Crowd Behavior Recognition

We formalize a series of processing steps which can be used to infer crowd behav-ior from location-based wireless information. In this section we build a mathematicalmodel for recognizing the crowd behavior of population.

First, based on the processed dataset, we confirm whether the crowd behavior of in-dividuals can be identified. Fig. 2 shows the results of data clustering with Friendshipand Location id attributes, respectively (these two attributes are evaluation classesfor clustering), using Expectation-Maximization (EM) algorithm [8]. From Fig. 2, the

6 Yuanfang Chen et al.

crowd behavior has been recognized using the processed dataset (in our study a crowdis defined in Definition 1). Moreover, we can find that different evaluation classes havedifferent clustering accuracy levels, so the impacts of different attributes on the crowdbehavior are different. Fig. 3 shows the number of users for each cluster correspondingto Fig. 2 (Fig. 2 only shows the users who are “1-hop” friends, but in Fig. 3, the userswho have multihop friendship are also counted for each cluster).

Definition 1 (A Crowd). A group of individuals at the same physical location (therange radius r = 10m) at the same time.

1071

1097245914

7201

44126

8724

10650

45913

646

6908 71556933

7170

2908

2897

629

179

174

4

2832

1076

21818

82402808

3010752

1

156

297511

2947

0

10986

2894

2850

6208

15542

167

1065813501

10343

10717 6782

6677

2953

2822

1386

680

1437231

1053610767

7705

10436

890710669

10668

71152834

146

9569

10612

2844

2821

217

8467

2

26792

1875

1037410350

8869

7798

3173

1077

651

28198145 8924 13265

28337620

19113

10988

2806

7715

15313

7790

10426

2920

620

2912

11746

7602

15377

3080 10632

10653

159

336

1453

35288

1041113252

7517

318911029

314615565

15385

1515

3014229710974

10554

965878

11564956634

21121

6606

7537

13361

306346708869613560

64

2874562

28539

3670

1701

24659

2799724601

3491337296

23841

10729

6825

25

3668 66921164611204

7546 2949

2969 133233058

29912

12873

53458

30058

3680

2910

55

382037436

23908

10622

2918

18601075

3160

7681

28678

10967

7794

6904

106147281

2987

291113

46488

3083

23796

1307

22518

407995604

6625

1499

9406

7162

11016

11926 7103

15815

24597

2892

1446

23541954

177712574 9403

6863

679

13653

6614

698843796

15558 11504

9130

36188

2951

2929

1308279

19389

135489971

30705

39405

10714

106716600

7376

292423838

16451

6842

11519

28059

24604

301

108097843

33361

2818613535

7755

7647

3650

3206

7783

232228750

15820

7084 1867

762121629

232

1074310353

132

15350

1582920355

1556956071

4612

3233

2264

13479

12908470

683615844

419

263

1908419731984

119017590

7594

6589

6587259

219

1594529903

14814

399994078435833

1944

9

11003

1880

7847

16447

56344

47768

7837

9592

95941985

875

1520 1693

20477

16620

21696

15375

317023

17315

40002 6213

1940

832

3302

11014

1779513579

112999977 1914

1285

987

627

8902

1697

11663

11528

9416

6916

10959

317

2855 3109

10559

10685

28451876

23

2

9417

10362 15852

2865 2839

28887

385

6902

154

26

65992861

7528

3009

1033328888

10966

2993

1419

124

42

10354

109804

81256604

5404

3313

2880

2818

267 7

2182213503

7518

12985

10375

104232902

28257558

652

8889

026494

63411282

2896

103511935

109638946

10400

6915

3332

241

2954

10380

2893

9847

7527

10990

15727

55514

6980

6638

2905

2176

23544

35129

18848

9593 15557 7251

16448

2847

253

157

779311261

28518

57713 7619

2134

1771

293

144

13834

22271

11609

11824

10336

9985

99167526

6948

15335

35562

26137

2640537001

1126724103

16709

5628

2206

9918

10801

16450

8248

1859

310

7848

513896841

5351516696

754217045

35

775637380

25003

15262

21695

6913

16529

5655

1095

66213419

6956

2773093

6955 11836

35128

8249 287766738

2873632628 623

209

1491413

1142

15

26320

10382

10590

11264

1864

26214

6249

11816

11260

10875

6588

62

1923

7529

698

(a) Clustering result withFriendship attribute.

13419

8145

26792

8249

7715

6677

2951

2894

2821179

10986

26137

89247705

104112833

167

143

146

15377

13501

10653

7602

6915

15852662

277

5

35833

6614

72312920

652

10426

10374

10671

10729

10436

7798

10668

15313

336

35288

629

6208

13265

10988

6933

10767

10423

10658

7155 5404

10400

33322908

10343

2850 1386

3080

2822

14537620

262

159

2905

2912

106693093620

8240

6956

10967

7790

1562806

2819

2897109632954

279

174

1

2825

7170

8467

10336

2844

2902

132

2947

651

283215542

6913

678213252

0

68010350

235

103541875

1864

28181076

6588 241

7755

3173

55

2181827997

7

07528

6904

28888

30

10375

2808

7115 26320

1876

1446

1419

1071646

9

10380

7756

21822

11014

3109

2839

124

26

2873

6863

11504

10362

12985

15385

8902

6587

46708

40002

3146

28059

24601

3058987

627

8946

1063210875

10351

8889 824814

15350

16620

1940

170231914

1701

36806213

28678

5634415829

11646

263

1190111519

15727

216961156

7103

13

11003

7681

1059011746

1584429912

7847

11609913010959

2910

2892

16450107171142

22271

24103

1582040784

28518

35129

2834

875012873

3

6842

761910974

1152812574

6841

13479

28186

407999847

2134

1944708423222

11663 28539

10714

9416

135793014

1307

419

219

53458

13653

19113 10622

72816916

2911

2847

1308

764711926

3336123838

1183641954

16448

9406

8470

7542

4

6948

6249

5628

1859

878

62

15557

10536

16529

6738

118241923

2206

1095

632

23841

7518

11261 45913

45914 16696

9971

7529 310

149

15

2949

6692

8869

13548

144

875

6709

9977

6955

169713834

198435128

1075

2924

2176

26494

66042874

2888710333

10650

1880

42

23

2896

679623

112671771293

1693

21695

3313

2929

209

10685

1935

9417

3009

2880

385

2162911282

2865

2993267

11

4

25

9569

2953

10382

7558

10612

7517

10559

6902

11299

154

13503

10809

634

11016

7201

7527

284510972

4412664

7526

11260

10966

109801860

34

2861

6599

6980

8125

8724

2975

10752

2855

1777

14132

30705

3160

19389

47768

3302

17795

7621

35562 1520317

9918

56071

24659

2969

13560

46488

3729625003

3206

7794

11264959213361

7376 24597

298714814

159456606

3233

1985

2893

7537149929903

9985 3743615375

43796

6282918

10614

15335

17315

232

217

7783

203557837

16451

7162

2264

22518

21121

869655514

10801

51389 725175466600

83253515

23544

20477

39405

5655

1867

1077

9594

155656638164471120418848

3820

1515

259

35

7590

36188

373801157

110297594

419737843

9593

46122297

1332323796

991657713

7793

68366589

1285 562

34913158153670

9403

1290965

301152623063 5604

3083

300587848

13535

39999

69811816

6825

239086988

1099015558

6908

15569

3189

24604370018907

10743

66253650

1908

253

3668

4956617045

26405

28776

10353

10554

(b) Clustering result withLocation id attribute.

Fig. 2. For clearness, we only show one-month clustering result using Friendship andLocation id attributes as the classes of clustering evaluation, respectively. 5 clusters can be foundfor two clustering processes, and we use different colours to distinguish different clusters. Notethat the shape for the center node of any cluster is square.

The characteristics of the crowd behavior of each single person can thus be in-ferred from his/her check-in records. We refer to this as their “individual behavior”.This shows which individuals participate in a specific crowd. From the EM algorithm,a given record belongs to each cluster with certain probabilities. Moreover, the likeli-hood is a measurement of “how good” a clustering process is and it is increased at eachiteration of the EM algorithm. It is worth mentioning that the higher the likelihood, thebetter the model fits the data.

Second, the clustering process (crowd behavior recognition model) is as follows.We define two parameters: (i) the user u’s check-in data S u which is a sequence ofactivity observations about u; (ii) a set of unknown values θ (i.e., the serial numbers ofclusters). These two parameters are used along with a Maximum Likelihood Estimation(MLE): L(θ; S u) = p(S u|θ). Our purpose is to seek the MLE of marginal likelihood. Inother words, we need to find the most probable θ which the user u belongs. The EMalgorithm iteratively applies the following two steps to achieve our purpose:

1. Expectation step (E step): calculate the expected value of the log-likelihood func-tion under the current established clusters (θ(t)):

Q(θ|θ(t)) = ES u,θ(t) [log L(θ; S u)];

Crowd Behavior 7

1 2 3 4 50

500

1000

1500

2000

2500

3000

3500

Cluster number

Num

ber o

f use

rs

(a) Number of users for eachcluster using Friendship at-tribute as the class of clusteringevaluation.

1 2 3 4 50

500

1000

1500

2000

2500

3000

3500

4000

Cluster number

Num

ber o

f use

rs

(b) Number of users for eachcluster using Location id at-tribute as the class of clusteringevaluation.

Fig. 3. Number of users for each cluster.

2. Maximization step (M step): find the appropriate value of parameter θ, which max-imizes this quantity:

θmle = arg maxθ

Q(θ|θ(t)).

MLE estimates θ by finding a value of θ that maximizes Q(θ|θ(t)), and the estimationresult can be flagged as: θmle.

Further, based on our recognition model, we add the spatio-temporal pattern intothe crowd behavior (clustering). We use a triple q = (θ, pi, ti) to replace θ. Then theexpectation step becomes:

1. Expectation step:Q(q|q(ti)) = ES u,q(ti ) [log L(q; S u)],

where pi is the position characteristic of location measurement mi, and ti is times-tamp of mi. Moreover, q(ti) is a set of current established clusters with their locationsand timestamps;

2. Maximization step: choose q to maximize Q(.),

qmle = arg maxq

Q(q|q(ti)).

Finally, in order to preserve the model integrity, the recognition accuracy must bemeasured. In our crowd behavior recognition model, the value of the log-likelihood canbe used to measure the accuracy. For instance, using the Friendship attribute as theevaluation class, based on one-month data, the log likelihood of crowd identification is:−16.42186, and using the Location id attribute as the evaluation class, the log likeli-hood is: −16.87742. Their accuracy is different, and the Friendship attribute is moreeffective for improving the recognition ability of the model.

8 Yuanfang Chen et al.

5 Impact of Attributes on Crowd Behavior

Most previous studies of crowd behavior only consider interactions among isolated in-dividuals, and assume that the motion of individuals is random. The work in [22] af-firms that the walking behavior of pedestrians is affected by social relationships, suchas friends, couples, or families walking together. The results presented in Fig. 2, showthat some population attributes impact the crowd motion of humans, and these impactsare different for diverse attributes. In this section, we go deeper into the impact of theseeffects by analyzing how friendship and location affect crowd behavior.

5.1 Impact of Friendship on Crowd Behavior

In this section, we analyze the impact of social relationship (friendship) on the complexdynamics of crowd behavior. For this, we use the empirical data of the motion of indi-viduals by means of check-in recordings of public areas. Observations are made undervarying-density collections of population.

0 1 2 3 4 5 6x 104

0

10

20

30

40

50

60

70

UserID

Deg

ree

(k)

User−Degree Distribution

January dataFebruary data

(a) Friendship degree of each us-er.

0 10 20 30 40 50 60 700

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Degree (k)

P(k)

Degree Probability Distribution

February dataJanuary data

(b) Probability distribution of var-ious degrees.

Fig. 4. Friendship degree of each user and probability distribution of various degrees (Januaryand February, 2010 datasets).

First of all, we analyze the friendship of a group of people. Fig. 4 shows the friend-ship degree of each user and the probability distribution of various degrees (based onthe data of January and February, 2010). We can find that friendship exists in almost allobserved users. It means that friendship is an important influencing factor for humanbehavior. It is worth noting that a Poisson distribution is met for the friendship degreeprobability distribution of all observed users. Moreover, from Fig. 4(b), friendship de-gree is less than 10 for almost 90% users. So the crowding propensity of an individualis not primarily oriented by friendship.

Secondly, several crowds exist in the observed group of people and we investigatethe friendship degree probability distribution of each crowd. Fig. 5 shows the friendshipdegree distributions of 5 clusters which are based on January, 2010 dataset. For some

Crowd Behavior 9

crowds, friendship degrees show approximative scale-free power-law distributions, e.g.,crowds 0, 1, 2, and 3.

100 101 102100

101

102

Degree (k)

Num

ber o

f use

rs 0

100 101 102100

102

104

Degree (k)

Num

ber o

f use

rs 1

100 101 102100

101

102

Degree (k)

Num

ber o

f use

rs 2

100 101 102100

102

104

Degree (k)

Num

ber o

f use

rs 3

100 101 102100

101

102

Degree (k)

Num

ber o

f use

rs 4

Fig. 5. Relationship between friendship degree and number of users for each crowd (January,2010 dataset).

0 1 2 3 40

1

2

3

4

5

6 x 104

Cluster

Frie

ndID

Fig. 6. IDs of users (there is friendship among these users) for each crowd.

Finally, how many users have friendship between each other for each cluster? Fig. 6exposes an interesting phenomenon: even though the average friendship degree of clus-ter 2 is larger than that of cluster 3 (shown in Fig. 5), the number of users who havefriendship in cluster 2 is less compared to cluster 3. This means that there is no relationbetween the number of users (who have friendship) and friendship degree in a crowd.

5.2 Impact of Location on Crowd Behavior

With the increasing ubiquity of location sensing included in mobile devices, we realizethe arising opportunity for human crowd analysis through mobile wireless informationfrom the real world context. Thereby, the spatial locations of individuals become a

10 Yuanfang Chen et al.

kind of important and available information for extending the analysis and modeling ofhuman crowd behavior to the physical world. Most prediction models of crowd eventsuse some kind of location information. In [4], Calabrese et al. believe that the attendeesof crowd events are related to areas: “Sport events such as baseball games attract aboutdouble the number of people which normally live in the Fenway Park area. Moreover,those events seem to be predominantly attended by people living in the surrounding ofthe baseball stadium, as well as the south Boston area”. From our investigation, we canfind the impact of location on crowd behavior. Fig. 7 shows location id distribution ofusers.

Loca

tion_

id

UserID0 60000

8000

4000

6000

2000

3000015000 450000

(a) Location distribution of clus-tered users.

Check-in time

Loca

tion_

id

00

10000

5000

2500

7500

70003500

(b) Along with the change of check-in time, the location distribution ofclustered users.

Fig. 7. Relationship among users, check-in time, clusters and locations (different colours denotedifferent clusters).

Analyzing Fig. 7, we can see that users are obviously clustered, and an approximatediagonal line divides the User-Location (see Fig. 7(a)) and Check-in time-Location (seeFig. 7(b)) space. Moreover, we can observe: the range of activity for any user is limited;and duration time is different for different crowds in different locations. We can assertthat human crowd behavior is centralized; namely “hot spot” exists. We use this char-acter of crowd behavior to do centralized population monitoring or urban public facilitydeployment.

Cluster

Loca

tion_

id

Fig. 8. Location distribution of users in each cluster.

Figure 8 shows the location distribution of users for each crowd. We can find thatthe distances between users are different (the location distribution of users is uneven) in

Crowd Behavior 11

any crowd. For instance, in cluster 3, there are two small subclusters. We believe thatthe small-world phenomenon exists in crowd behavior.

6 Conclusion

The analysis of crowd behavior can give us an idea of how we behave when we arepart of a group. Different actions can be taken by individuals when being surroundedby others. In this paper, we have focused on analyzing crowd behavior, consideringcheck-in data from location-aware mobile social networks.

There are several parameters that affect the forming way of a crowd. Social relation-ships, e.g., friendship, as well as current locations, can impact the way that a crowd isformed. Understanding how these parameters determine crowd formation and evolutionis the first step prior to the creation of predictive approaches. After grasping the rela-tionship between these parameters and crowd behavior, our ongoing research focuseson the prediction model of crowd behavior. This model will be useful for preventingdisasters from the pushing of crowds, facilitating efficient massive event planning, oreven for traffic management.

Acknowledgment

This work was supported by the EU ITEA 2 Project 11020, “Social Internet of Things-Apps by and for the Crowd” (SITAC).

References

1. Antonini, G., Bierlaire, M., Weber, M.: Discrete choice models of pedestrian walking behav-ior. Transportation Research Part B: Methodological 40(8), 667–687 (2006)

2. Bandini, S., Manzoni, S., Vizzari, G.: Crowd behaviour modeling: from cellular automatato multi-agent systems. Multi-Agent Systems: Simulation and Applications pp. 204–230(2009)

3. Bellomo, N.: Modeling crowds and swarms: congested and panic flows. Modeling ComplexLiving Systems pp. 169–188 (2008)

4. Calabrese, F., Pereira, F., Di Lorenzo, G., Liu, L., Ratti, C.: The geography of taste: analyzingcell-phone mobility and social events. Pervasive Computing 6030, 22–37 (2010)

5. Cho, E., Myers, S.A., Leskovec, J.: Friendship and mobility: user movement in location-based social networks. In: Proceedings of the 17th SIGKDD International Conference onKnowledge Discovery and Data Mining, pp. 1082–1090. ACM (2011)

6. Cook, D.J., Das, S.K.: Pervasive computing at scale: transforming the state of the art. Perva-sive and Mobile Computing 8(1), 22–35 (2012)

7. Davis, L.: Predicting travel time to limit congestion at a highway bottleneck. Physica A:Statistical Mechanics and its Applications 389(17), 3588–3599 (2010)

8. Do, C.B., Batzoglou, S.: What is the expectation maximization algorithm? Nature biotech-nology 26(8), 897–899 (2008)

12 Yuanfang Chen et al.

9. Eagle, N., Pentland, A.S., Lazer, D.: Inferring friendship network structure by using mobilephone data. Proceedings of the National Academy of Sciences 106(36), 15,274–15,278(2009)

10. Ge, W., Collins, R.T., Ruback, R.B.: Vision-based analysis of small groups in pedestriancrowds. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(5), 1003–1016(2012)

11. Gonzalez, M.C., Hidalgo, C.A., Barabasi, A.L.: Understanding individual human mobilitypatterns. Nature 453(7196), 779–782 (2008)

12. Griesser, M., Ma, Q., Webber, S., Bowgen, K., Sumpter, D.J.: Understanding animal group-size distributions. PloS one 6(8), e23,438:1–9 (2011)

13. Helbing, D., Molnar, P.: Social force model for pedestrian dynamics. Physical review E51(5), 4282–4286 (1995)

14. Jeung, H., Liu, Q., Shen, H.T., Zhou, X.: A hybrid prediction model for moving objects.In: Proceedings of the 24th International Conference on Data Engineering, pp. 70–79. IEEE(2008)

15. Jiao, Y., Liu, Y., Wang, J., Wang, J.: Model for human dynamics based on habit. ChineseScience Bulletin 55(24), 2744–2749 (2010)

16. Ko, M.H., West, G., Venkatesh, S., Kumar, M.: Online context recognition in multisensorsystems using dynamic time warping. In: Intelligent Sensors, Sensor Networks and Infor-mation Processing Conference, pp. 283–288. IEEE (2005)

17. Lazer, D., Pentland, A., Adamic, L., Aral, S., Barabasi, A.L., Brewer, D., Christakis, N., Con-tractor, N., Fowler, J., Gutmann, M., Jebara, T., King, G., Macy, M., Roy, D., Van Alstyne,M.: Computational social science. Science 323(5915), 721–723 (2009)

18. Mitchell, T.M.: Mining our reality. Science 326(5960), 1644–1645 (2009)19. Monreale, A., Pinelli, F., Trasarti, R., Giannotti, F.: Wherenext: a location predictor on tra-

jectory pattern mining. In: Proceedings of the 15th ACM SIGKDD international conferenceon Knowledge discovery and data mining, pp. 637–646. ACM (2009)

20. Moussaıd, M., Helbing, D., Garnier, S., Johansson, A., Combe, M., Theraulaz, G.: Experi-mental study of the behavioural mechanisms underlying self-organization in human crowds.Proceedings of the Royal Society B: Biological Sciences 276(1668), 2755–2762 (2009)

21. Moussaid, M., Helbing, D., Theraulaz, G.: How simple rules determine pedestrian behaviorand crowd disasters. Proceedings of the National Academy of Sciences (PNAS) 108(17),6884–6888 (2011)

22. Moussaıd, M., Perozo, N., Garnier, S., Helbing, D., Theraulaz, G.: The walking behaviourof pedestrian social groups and its impact on crowd dynamics. PloS one 5(4), e10,047:1–7(2010)

23. Paul, U., Subramanian, A.P., Buddhikot, M.M., Das, S.R.: Understanding traffic dynamics incellular data networks. In: Proceedings of the 30th International Conference on ComputerCommunications, pp. 882–890. IEEE (2011)

24. Pentland, A., Choudhury, T., Eagle, N., Singh, P.: Human dynamics: computation for orga-nizations. Pattern Recognition Letters 26(4), 503–511 (2005)

25. Pentland, A.S., Pentland, S.: Honest signals: how they shape our world. MIT press (2008)26. Tang, L., Liu, H.: Toward predicting collective behavior via social dimension extraction.

Intelligent Systems 25(4), 19–25 (2010)27. Trasarti, R., Pinelli, F., Nanni, M., Giannotti, F.: Mining mobility user profiles for car pool-

ing. In: Proceedings of the 17th SIGKDD International Conference on Knowledge Discoveryand Data Mining, pp. 1190–1198. ACM (2011)

28. Zhang, X., Weng, W., Yuan, H., Chen, J.: Empirical study on unidirectional dense crowdduring a real mass event. Physica A: Statistical Mechanics and its Applications 392(12),2781–2791 (2013)