understanding the behavioral differences between american...

BIG DATA MINING AND ANALYTICSISSN ll1007-0214 0?/?? pp???–???DOI: 10.26599/BDMA.2018.9020024Volume 1, Number 1, Septembelr 2018

Understanding the Behavioral Differences Between American andGerman Users: A Data-Driven Study

Chenxi Yang, Yang Chen∗, Qingyuan Gong, Xinlei He,Yu Xiao, Yuhuan Huang, Xiaoming Fu

Abstract: Given that the USA and Germany are the most populous countries in North America and Western

Europe, understanding the behavioral differences between American and German users of online social networks

is essential. In this work, we conduct a data-driven study based on the Yelp Open Dataset. We demonstrate

the behavioral characteristics of both American and German users from different aspects, i.e., social connectivity,

review styles and spatiotemporal patterns. In addition, we construct a classification model to accurately recognize

American and German users according to the behavioral data. Our model achieves high classification performance

with an F1-score of 0.891 and AUC of 0.949.

Key words: Behavioral Difference; Online Social Networks; Yelp; Machine Learning

1 Introduction

Cultural differences are a core question of interestamong sociologists. Over the past decades, thecultural differences, reasons behind these differences,and phenomena that these differences reflect, includingcollectivism, individualism and social sustainability,have been intensively studied [1–3]. In these

•Chenxi Yang, Yang Chen, Qingyuan Gong and Xinlei Heare with the School of Computer Science, Fudan University,China, and the Engineering Research Center of Cyber SecurityAuditing and Monitoring, Ministry of Education, China. E-mail: [email protected]

•Yu Xiao is with the Department of Communicationsand Networking, Aalto University, Finland. E-mail:[email protected]

•Yuhuan Huang is with the Faculty of European Languages andCultures, Guangdong University of Foreign Studies, China. E-mail: [email protected]

•Xiaoming Fu is with the Institute of Computer Science,University of Gottingen, Germany. E-mail: [email protected]

∗To whom correspondence should be addressed.Manuscript received: 2018-03-05; accepted: 2018-03-20

studies, data for understanding users’ culture-relatedbehaviors is obtained using traditional methods, such asquestionnaires, video documentation and other personalinterview methods.

Compared with rather static online textual data(which remains untouched once published) suchas news or web pages, situation-aware interactiveinformation like user reviews and comments providesmore daily life-related and accessible (typically viatext) opinions and thoughts. Nowadays, onlinesocial networks (OSNs) [4] have witnessed rapidgrowth, attracting billions of users worldwide. Peoplecontribute to profiles, social activities, and life trackson OSNs, and deep cultural impacts exist among thesebehaviors. Krasnova et al. [5] conducted a surveyon Facebook to explore the differences in individualwillingness to self-disclosure between American andGerman users via a questionnaire to Facebook users.They concluded that American users are more activeon Facebook and have higher privacy concerns thanGermans. To our knowledge, [5] is the first work thathas analyzed users’ online behavior from a culturalperspective. However, given that most of the cultural

2 Big Data Mining and Analytics, September 2018, 1(3): 000-000

phenomena evolved for many years and developed fromgeneration to generation, the scale of the online survey-based research is, although larger than the survey inthe real world, still not large enough to form a culturalimpact. Further, data comprehensiveness is of greatconsequence to cultural analysis. Apart from the answerto the questions in the survey or text posted by users, themovement pattern and points-of-interest (POIs), whichare closely related to the cultural impact on a user, alsomatter.

To this end, location-based social networks (LBSNs),such as Yelp [6], Foursquare/Swarm [7–9], Momo [10],Skout [11] and Dianping [12, 13], which allow users toundertake location-centric activities in addition to socialinteractions, offer a viable data source for such culturalstudies. These LBSN platforms record the activitydata of massive users and provide researchers a greatopportunity to compare human behavior from bothspatiotemporal and social networking perspectives.

The USA and Germany, which have the largestpopulations in North America and Western Europe,respectively, are two important culture clusters in theworld. They have different languages, traditions, andgeographical conditions but also share the same Anglo-Saxon origins. First, Germans care more about socialstability than Americans [14], whereas Americans tendto become outstanding individuals because of their eliteculture [1]. Second, in terms of collectivism, Germansprefer risk taking and uncertainty avoidance comparedwith Americans [15]. However, the differences betweenGerman and American users in terms of behavioralpatterns are rarely studied. We aim to know whether thebehavior of users on LBSNs is consistent with previouscross-cultural research results [5,16] and if not, identifywhich aspects of behavior have changed online.

In this paper, we use data from a representativeLBSN, Yelp, as basis for our study. We conduct adata-driven study based on the Yelp Open Dataset∗.Yelp can help people discover local businesses, a.k.a,POIs, such as restaurants, hospitals, or spas. Itallows users to publish reviews or conduct check-ins in selected businesses. Yelp users completedmore than 142 million reviews by the end of Q32017†. Moreover, Yelp serves more like an “urbanguide”, a review platform with location informationand category preference, rather than a channel thatonly helps you make friends. This platform partially∗https://www.yelp.com/dataset†https://www.yelp.com/about

reflects various aspects of people’s daily life, whichallows the inference of the entire profile and clear socialengagement of a user. Given its popularity and richuser-generated content, Yelp is selected for our userbehavior study. Compared with [5], social connectivity,spatiotemporal patterns and writing styles are combinedbased on the data from a much larger scale of usersin our work. We select the USA and Germany as theexamples to understand online behavior from a culturalperspective and make comparisons between these tworepresentative cultural clusters in North America andWestern Europe respectively. Our key contributions aresummarized as follows.

• We provide a comprehensive statistical anddemographic analysis of American and Germanusers’ behavior based on the Yelp Open Dataset,and compare the results in a comprehensivemanner. We find that American users aremore influential on Yelp than German users,and their friends are scattered in more cities.Our spatiotemporal analysis shows that Germanusers have a clearer line between daytime andnightlife than American users. On the basisof our analysis of review texts, we also provethat collectivism is important for German users,whereas individualism is a priority for Americanusers.

• In this paper we verify the feasibility of applyingbig data analysis in the context of cultural behavior.In particular, we employ our analysis of users’online behavior to construct a classification modelthat can accurately detect whether a user is fromthe USA or Germany. With this classificationmodel, we achieve an F1-score of 0.891 and AUCof 0.949 for detecting whether a Yelp user is fromthe USA or Germany, which serves as a strongbuttress of our analysis and feature selection. Wefind that writing style- and social graph-relatedfeatures are the most distinguishing features todifferentiate American and German users.

The rest of the paper is structured as follows. Wefirst introduce Yelp and the dataset used for our studyin Section 2. In Section 3, we analyze the data for bothuser groups and businesses on Yelp from a cultural viewof USA and Germany and identify the features that arestrongly related to cultural diversity. We then providecomprehensive evaluations on our classification model

Yang et al.: Understanding the Behavioral Differences Between American and German Users: A Data-Driven Study 3

using various supervised machine learning algorithms,including the importance of different feature sets, inSection 4. We review the related work in Section 5 andconclude this work in Section 6.

2 Background and Dataset

In this section, we provide an overview of Yelp andthen introduce the dataset used in this study.

2.1 Background of Yelp

Founded in 2004, Yelp.com has become one ofthe world’s largest online “urban guide” and businessreview sites [17]. On Yelp, users can write reviews,upload photos, conduct check-ins, and rate theirexperiences at different types of businesses such asrestaurants and hotels. Yelp allows a user to conducta check-in at a business only when the user is closeenough to the business. In addition, users are supposedto give a review of a business several days after theirvisit. Yelp covers 21 main categories and over 1,200sub-categories of businesses ‡. It provides a platformfor users to express their preferences over differentbusiness categories. Meanwhile, it serves as a socialnetworking platform. Users can make friends with otherusers who show interests to similar business categories.Together with reviews and check-ins, the data aboutusers’ preferences and friends reflect user behavior indaily life in an informative way.

2.2 Dataset Description

We study the Yelp Open Dataset, which was used inthe Yelp Dataset challenge. The dataset covers over4,700,000 reviews, 156,000 businesses, and 1,100,000users. Each review contains text and/or rating attributes.

The dataset is composed of 11 tables. For each user,we can obtain his/her friends, the year when the userstarted using Yelp, average number of stars, and othercomprehensive assessments of his/her reviews and tips.For each business, its location, category, reviews, tips,and check-ins are all available. Regarding each user’shome city, we assume that the user belongs to the citywhere he or she reviews most, defining the city asthe “home city” and getting the country informationaccordingly. Regarding users’ home countries, theUSA, Germany, Canada and UK are the four maincountries.

‡www.yelp.com/developers/documentationFusion

3 Data Analysis

In this section, we study the behavior of Americanand German users using the Yelp Open Datasetdescribed in Section 2. Our goal is to betterunderstand the differences and similarities of Americanand German users in terms of location distributionof friends, social graph characteristics, writing styleof reviews, preference for business categories, ratingpreferences, and temporal patterns of check-ins. Thissection is divided into three subsections, i.e., socialgraph, reviews, and check-ins.

3.1 Analysis of the Social Graph

To understand the social behavior of Americanand German users on Yelp, we use the StanfordNetwork Analysis Platform (SNAP) [18] for socialgraph analysis. SNAP is a general purpose networkanalysis and graph mining library written in C++.Using SNAP, we then analyze some representativenetwork metrics, i.e., degree, clustering coefficient(CC), PageRank and connected components in Yelp’ssocial network G.

3.1.1 Analysis of the Social Graph as a WholeThe Yelp’s friendship network, G, has 8,981,389

nodes and 35,444,850 edges. Fig. 1(a) shows thecumulative distribution function (CDF) of the degreesin G. The degree of a node represents the numberof edges connected to the node. A higher degree inG means that the user has more friends. No nodeshave zero degrees, which means that any user on Yelphas at least one friend. The average degree of Gis 7.0. Compared with many other OSNs, 7.0 is asmall average degree. Given that Yelp is not a websitededicated to social networking, users on Yelp are notchasing after a large number of friends. Therefore, theconnection on Yelp is looser than that in most OSNs.

The CC measures the cliquishness of a typicalfriendship circle. A higher average CC indicates thatit is more likely for nodes to form tightly knit groups.Fig. 1(b) is the CDF of CC of the nodes in G. Over70% of Yelp users have a CC of zero and the friendshipnetworks’ average CC is 0.055, which demonstrates aweak connection between Yelp users. This result is dueto users that employ Yelp as a guide service rather thana networking site.

PageRank is a metric that quantifies the importanceof different nodes in a network [19], which has beenapplied by the Google search engine to rank websites.


Nation Avg. CC Var. CC Avg. Degree Var. Degree Avg. PageRank Var. PageRankUSA 0.054 0.030 7.0 2379.0 1.180× 10−7 4.405× 10−8

Germany 0.041 0.032 2.0 909.0 2.120× 10−5 8.360× 10−8

Table 1 Graph Attributes

The PageRank value of any node of a network isbetween 0 and 1. A higher PageRank value suggeststhat the corresponding node is more important in thisnetwork. In Fig. 1(c), the CDF of PageRank displayssimilar results.

Fig. 1(d) shows the sizes of the top 10 connectedcomponents. The connected component is a subgraphin which any two nodes are connected to each otherby paths, and this subgraph connected to no additionalnodes in the supergraph. A total of 18,512 connectedcomponents exist in network G. The largest connectedcomponent has 8,938,630 users, covering 99% users inthe Yelp Open Dataset. The second largest connectedcomponents have 15 nodes and the third to sixth have14 nodes each. We also calculate the distribution ofthe shortest path lengths of all node pairs in the largestconnect component (LCC), as displayed in Fig. 1(e).The average shortest path length is 4.93, and the averageCC of G is 0.055.

We further study the density of the “core” of the Yelpfriendship network G. From 0.01% to 10%, we removesome nodes with the highest degrees from the networkand analyze the remaining nodes. As in Fig. 1(f), wegroup the remaining nodes into three categories, i.e., theLCC, singletons, and middle region. We find that thenumber of singletons already surpasses that of the nodesstill in LCC after we remove 5% of the nodes with thehighest degree. Therefore, Yelp’s friendship network isnot as strongly connected as other mainstream OSNs,such as Renren [20] and Cyworld [21].

3.1.2 Comparison Between American and GermanUsers

Table 1 shows the mean and variance of CC ofAmerican and German users’ social graph, whichindicates that the mean of CC of American users isslightly higher than that of German users. We alsocompute the mean and variance of degrees in theAmerican and German users’ social graph. The meanand variance of degree of American users are bothlarger than those of German users. In other words, someAmerican users make a lot more friends on Yelp thanGerman users.

In Table 1, both the mean and variance of PageRank

values of German users are larger than those ofAmerican users. The PageRank values are used todefine the top 0.1% users of the whole social graphas “influential users”. We use one set, P, to representthe influential users of the entire graph. We also findthat 0.31% of the American users belong to P, whereasfor German users, that number is 0.18%. These resultsillustrate that both the USA and Germany have moreinfluential users than the average level of the entirenetwork of Yelp. Moreover, the proportion of influentialusers is higher in American users than in German users.

Fig. 2 reflects the location distributions of friends ofAmerican and German users. The x-axis represents thenumber of cities where the users’ friends are spread.The y-axis indicates the percentage of users who havefriends in a certain number of cities. As shown in Fig. 2,the share of German users who only have Yelp friendsin one or two cities is greater than that of Americanusers. When the number of cities is three or higher, theshare is about 15% larger in American users. We findthat German users’ friends are gathered in few cities,whereas the location distribution of American users’friends is slightly wider than that of German users’.

3.2 Analysis of Reviews

Apart from the analysis of social graph in Section 3.1,we also group the reviews in accordance with thecountry where the businesses and the users are located.In Section 3.2.1, the text attribute of review serves asa significant part when analyzing the character trait.We then give the category preference of Americanand German users in Section 3.2.2. Subsequently, wepresent the visit and rating preference in Section 3.2.3.

3.2.1 Writing Style AnalysisLinguistic Inquiry and Word Count (LIWC) [22] is

used to fully understand the text of reviews. It hasbeen widely used in computerized text analysis to learnhow the words we use in everyday language reveal ourthoughts, feelings, personality, and motivations. In ourstudy, given that the text of reviews comes from twolanguages, namely, English and German, we also usethe German LIWC2001 Dictionary [23].

As shown in Table 2, the numbers represent theoccurrence frequency of a specific set of words. In


0

20

40

60

80

100

0 20 40 60 80 100

Perc

enta

ge (

%)

Degree

(a) Degree

0

10

20

30

40

50

60

70

80

90

100

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Perc

enta

ge (

%)

Clustering Coefficient

(b) Clustering Coefficient

0

10

20

30

40

50

60

70

80

90

100

0.5 1 1.5 2 2.5 3 3.5 4 4.5 5x10-7

Perc

enta

ge (

%)

PageRank

(c) PageRank

100

101

102

103

104

105

106

107

108

1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th

Com

ponent S

ize

Top 10 Connected Component

(d) Connected Component Sizes

0

10

20

30

40

50

0 2 4 6 8 10 12 14 16 18 20

Perc

enta

ge (

%)

Path Length

(e) Path Length

0

10

20

30

40

50

60

70

80

90

100

0.01% 0.1% 1% 5% 10%

Perc

enta

ge (

%)

Fraction of Network Removed

SingletonsMiddle Region

LCC

(f) Robustness

Fig. 1 Analysis of the Friendship Network

Nation Affect Anger Tenta Certain Swear FriendsUSA 6.045 0.252 2.298 1.845 0.087 0.438

Germany 4.688 0.191 0.965 2.842 0.017 0.288

Table 2 Occurrence Frequency of Different Categories of Words in Reviews

Nation Category I We Leisure

USA

Beauty & Spas 7.02 0.53 1.16Health Medical 6.85 0.61 0.63Home Services 4.89 1.61 0.79

Nightlife 3.72 1.79 2.68Restaurant 4.08 1.49 1.58Shopping 5.47 0.88 1.44

Avg. 5.34 1.15 0.438

Germany

Beauty & Spas 4.17 0.27 1.06Health Medical 3.58 0.31 1.81Home Services 2.16 1.05 1.7

Nightlife 1.73 1.29 1.52Restaurant 1.78 1.34 1.42Shopping 3.04 0.46 1.15

Avg. 2.74 0.79 1.44

Table 3 Dimension Values of American and German Users

the first two columns, “Affect” and “Anger”, whichrepresent the affective process with the example words“happy”, “ugly”, and “bitter” and contents with “hate”,“kill”, and “pissed” [24] have something to do withindulging in anger. The average occurrence of“Affect” and “Anger” is higher among American usersthan among German users. As in [25], Americanstudents behave in a more affective way than German

students and this conclusion conforms to our resultsfrom emotions behind writing style. With regardto the “Tenta” (representing the tentative words) and“Certain” (representing the certainty), American usersprefer to write tentative-related words such as “maybe”,“perhaps” and “guess” [24], whereas German usersmention certainty-related words like “always” and“never” frequently. For “Swear”, including words like


0

10

20

30

40

50

1 2 3-5 6-10 >=11

Perc

en

tage (

%)

City Numbers

American usersGerman users

Fig. 2 Location Distribution of Friends

“fuck”, “damn” and “shit”, German users are less likelyto use swear words than their American counterpartswhen reviewing on Yelp. Words such as “buddy” and“neighbor”, which belong to the “friends” category,appear more frequently in American users’ reviews thanin German users’ reviews.

In Table 3, we also determine the differences inoccurrence frequency of “I”, “We” and “Leisure” inthe six main categories between American and Germanusers. The frequency of writing the reviews with thepronoun “I” by American users is twice than that ofGerman users. When talking with “We”, the mostpossible category American users are in is “Nightlife”,whereas it is “Restaurants” for German users. Webelieve the frequency of the usage of “We” can bepositively related to the high prevalence of peoplegoing to the particular category of businesses together.Therefore, “Nightlife” and “Restaurants” serve as thefavorite category of American and German users,respectively, when they go out for social gatherings.

3.2.2 Preference for Business CategoriesExperimentally, our results indicate the distinction

of category preference between American and Germanusers. We analyze the distribution of reviews in 10main categories. Fig. 3 displays the category pattern ofAmerican and German users, with the y-axis describingthe logarithmic coordinates of review percentage ofa certain category. For “Food”, “Nightlife” and“Shopping” categories, American and German usersshare similar preferences in general. For othercategories, visible differences are found. We find thatGerman users show great interests to “Restaurants” asGerman users’ percentage is 49%, which is much largerthan that of American users’ (41%). American users

0.1

1

10

100

Restaurants

Food

Nightlife

Shopping

Arts Entertainment

Beauty Spas

Health M

edical

Public Services

Hom

e Services

Automotive

Others

Pe

rce

nta

ge

(%

)


Fig. 3 Category Pattern

0

10

20

30

40

1-2 2-3 3-4 4-5 5

Perc

enta

ge (

%)

Stars


Fig. 4 Stars

visit businesses in the category of “Beauty & Spas”(including subcategories like “Barbers”, “Hair LossCenters” and “Day Spas”), “Health Medical”, “HomeServices” and “Automotive” (including subcategoriessuch as “Auto Detailing”, “Registration Services” and“Car Wash”) much more frequently than German users,except in “Public Services”.

3.2.3 Visit & Rating AnalysisFor the rating section, we also compute the star

preference for businesses of American and Germanusers. In the Yelp dataset, the average stars of Americanand German users are 3.73 and 3.78, respectively, withsimilar variance (1.18 and 1.01, respectively). Asshown in Fig. 4, the star does not follow a normaldistribution and most of the users have a rating of 3-5 in a five-point scale. German users prefer to gradebusinesses with a high point but not full. By contrast,


0

10

20

30

40

50

0 0.2 0.4 0.6 0.8 1

Perc

en

tage (

%)

Outer review percentage


Fig. 5 CDF of Outer Review Percentage of Users WhoseReview Count is Larger Than Zero

0

0.2

0.4

0.6

0.8

1

1.2

12 24 12 24 12 24 12 24 12 24 12 24 12 24

Here

now

Count P

erc

enta

ge(%

)

Mon Tue Wed Thu Fri Sat Sun

(a) USA

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

12 24 12 24 12 24 12 24 12 24 12 24 12 24

Here

now

Count P

erc

enta

ge(%

)

Mon Tue Wed Thu Fri Sat Sun

(b) Germany

Fig. 6 Checkin Patterns

American users are more likely to give 1 out of 5 or5 out of 5 compared with German users, who adopt amilder rating style.

Fig. 5 displays the CDF of the percentage of users’review out of their home city (we call it outer review

percentage in the rest of this paper) of users whosereview count is larger than zero. In general, Americanusers have a larger outer review percentage thanGerman users. Considering the friend distribution inSection 3.1.2, the location distributions of friends andreviews of American users are both wider, whereasthose of German users’ are both centralized to fewcities. Our results also exhibit the strong and positivecorrelations of visit and friendship distribution asdisplayed in [26].

3.3 Analysis of Check-ins

Temporal distribution of check-ins have been widelyused to characterize the behavior of LBSN users [27].We analyze the temporal patterns of check-ins atbusinesses in Fig. 6. The y-axis represents thepercentage of check-ins during each hour of a week.Fig. 6 illustrates that American and German usersboth conduct more check-ins between Monday andSaturday and much less on Sunday. Simultaneously,the differences in German users’ everyday noon peak(the first peak of a day) and night peak (the secondpeak of a day) can reach 6%, which is much moreexplicit than that of American users (0.03%). Giventhe exact time of everyday, we find that the lunch anddinner peaks of German users are around 11 a.m and 6p.m., respectively, whereas those of American users arearound 1 p.m. and 10 p.m., respectively.

4 Implementation of the CountryClassification Model

After comparing the behavior of social graph, reviewsand check-ins between American and German userson Yelp, we establish a broad sense of the behavioraldifferences between these two groups of users andattempt to understand the cultural influence behindthese behavioral differences. From an integrated view,we implement a model based on the features extractedfrom the analysis results in Section 3. This modelaims to predict a user’s home country and evaluatethe classification performance of our approach andother related proposals. In addition, we investigate theimportance of each feature to examine the extent thesevarious features can influence the behavior of Americanor German users on Yelp.

In this section, we include the implementation detailsand evaluation process of the classification model. First,we present a brief introduction to the tools and methodsused in our study. We then describe the training and


Social Graph Related

Writing Style Related

Visit & RatingRelated

CategoryRelated

Training Set

Test Set

1050 German Users1050 American Users

Machine Learning Algorithm Trained Model

66.67%

33.33%

Dataset TrainingFeature Sets Test Prediction

ParameterTuning

Fig. 7 Overview of the Classification Model

testing sets. Finally, we describe the importance ofeach feature set and assess the performance duringcountry classification. To better implement the machinelearning algorithms for the classification model, weadopt XGBoost [28] and Weka [29]. Specifically,XGBoost is a scalable machine learning system for treeboosting, which is widely used in machine learningcontests. Weka supports a collection of machinelearning algorithms for data mining tasks and it isimplemented in Java. The algorithms in Weka includebut are not limited to Random Forest (RF), SupportVector Machine (SVM), and C4.5 Decision Tree (J48).

To construct a training and validation dataset,we randomly select 700 American users and 700German users from the Yelp dataset. We adopt fourrepresentative metrics, namely, precision, recall, F1-score, and AUC, to evaluate the performance of ourclassification model. Precision refers to the fraction ofpredicted German users who are really German users.Recall represents the fraction of German users whoare accurately identified. F1-score is defined as theharmonic mean of precision and recall. AUC is shortfor “Area Under receiver operating characteristic (ROC)Curve”. Its value is equivalent to the probability that arandomly chosen positive example is ranked higher thana randomly chosen negative example [30]. We adoptseveral classic machine learning algorithms to train andvalidate our model using ten fold cross validation. Foreach algorithm, we apply grid search to find the “best”parameters, during which our goal is to achieve a highF1-score. After parameter tuning, we randomly selectanother 350 American users and 350 German usersfor testing and use the trained model to detect eachusers’ home country. Table 4 shows our classificationperformance.

4.1 Evaluating Classification Model as a Whole

As in Fig. 7, we first include all the 25 features asa whole for classification. We divide these features

into four sets as in Table 5. Seven features representthe review counts of seven main categories in business-related feature sets, four represent social graph-relatedfeature sets, 10 are in writing style-related feature sets,and four comprise visit and rating related-feature sets.Table 4 compares several supervised machine learningalgorithms including XGBoost, Random Forest (RF),C4.5 Decision Tree, SVM and BayesNet. In particular,we consider SVMp (with polynomial kernel). UsingMcNemar’s test [31] to compute statistical significance,we find that the prediction performance of nearlyevery two classifiers is significantly different (p<0.005,McNemar’s test). The only exception is when weconsider the RF and XGBoost classifiers, as thedifference between them is unremarkable (p>0.2,McNemar’s test). These two classifiers have muchbetter prediction performance than the other ones,implying that they both can be used for classifyingcultural belonging in practice. With the overallconsideration of F1-score and AUC, all the F1-scoresin our results from different algorithms are larger than0.850. Among them, RF performs the best with an F1-score of 0.891 and AUC of 0.949. We adopt RF inthe following subsection to compare the contributionsof different features.

4.2 Evaluating the Contribution of DifferentFeature Sets

To better understand the importance of differentkinds of features in the model, we list the χ2 (Chi-Square) statistics [32] of the top nine features. Asshown in Fig. 6, the most discriminating feature is“Pronoun”, which represents words like “I” and “You”.Meanwhile, the features from writing style analysissuch as “Preps (preposition)”, “Tentat (tentative)” and“Certain (certainty)” are more important than otherfeatures. The writing style-related feature set playsan important role in distinguishing between Americanand German users on Yelp. However, social graph-


Algorithm Parameter Precision Recall F1-score AUCRandom Forest maxDepth=13, numFeatures=5 0.893 0.891 0.891 0.949

XGBoost learning rate=0.01, min child weight=1, 0.891 0.890 0.890 0.901max depth=4, gamma=0.0, subsample=0.95,lamba=1, alpha=0, colsample bytree=0.75,boost=gbtree, objective=binary:logistic

Decision Tree (J48) confidenceFactor C=0.2, Instance/Leaf M=4 0.878 0.878 0.878 0.899SVMp Degree=3, Cost parameter=20.0 0.853 0.853 0.853 0.853

BayesNet default 0.862 0.862 0.862 0.923

Table 4 Comparison of Different Supervised Machine Learning Algorithms for the Classification Model

Category Description

Business

Restaurants and FoodNightlifeEvent Planning & ServicesHotel & TravelArt & EntertainmentBeauty & SpasHealth & Medical

Social Graph

Number of cities having friendsClustering coefficientDegreePageRank

Writing Style

Number of words per sentenceFrequency of occurrence of prepositionFrequency of occurrence of pronounFrequency of occurrence of “Anger”Frequency of occurrence of “Leisure”Frequency of occurrence of “Sad”Frequency of occurrence of tentative words like “maybe”, “perhaps”Frequency of occurrence of certain words like “always”, “never”Frequency of occurrence of “Friends”Frequency of occurrence of swear words

Visit & Rating

Number of reviewsNumber of visited citiesPercentage of visit out of home cityAverage star

Table 5 Subsets of Features of the Classification Model

Rank χ2 Feature Category1 969.876 Pronoun Writing Style2 650.939 Preps Writing Style3 366.716 Tentat Writing Style4 268.432 Certain Writing Style5 199.665 CC Social Graph6 99.615 Friend City Num Social Graph7 85.471 PageRank Social Graph8 73.282 Swear Writing Style9 60.701 Beauty & Spas Business

Table 6 Feature Importance: χ2 Analysis

related features such as “CC”, “Friend City Num” and“PageRank” are also of importance, ranking the fifth,

sixth and seventh, respectively, in Fig. 6. To understandmore details about other feature sets, we also evaluate


Category Presicion Recall F1-score AUCWriting Style 0.879 0.878 0.878 0.937Social Graph 0.741 0.741 0.741 0.823

Business 0.623 0.612 0.617 0.660Visit & Rating 0.602 0.603 0.601 0.619

Table 7 Contribution Analysis of Each Feature Subset in Classification Model

four feature sets independently and compare theirperformance. Results are shown in Table 7. Withthe 10 writing style-related features, we achieve anF1-score of 0.878 and AUC of 0.937. For the socialgraph-related feature set, F1-score is 0.741 and AUCis 0.823. For the other two feature sets, F1-scoreis 0.617 and 0.601, and AUC is 0.660 and 0.619 ofbusiness-related features and visit and rating-relatedfeatures, respectively. Corresponding to Chi-Squarestatistical analysis, the differences in writing style andsocial graph are the two most distinguishing attributesbetween users from USA and Germany on Yelp.Furthermore, preferences for business and visit andrating have a relatively slight effect on classification.

5 Related Work

Combining the features of social networks withgeographic information sharing, LBSNs are becomingincreasingly popular. Some recent works focused onexploiting online social interactions among individualsto explain social phenomena. For example, Topaet al. [33] generated a stationary distributionof unemployment which exhibited positive spatialcorrelations. Wal et al. [34] applied social networkanalysis in economic geography. For friendshipdistribution and user mobility, Liben-Nowell et al. [35]introduced a model capturing user behavior in real-world social networks and found that the probabilityof befriending a particular person and the number ofcloser people are inversely proportional. Cho et al. [26]discovered that long-distance travels are more likelyinfluenced by social network ties compared with short-ranged travels. Those studies provided strong evidencethat data on LBSNs can be utilized to analyze groupsof user behavior. However, no studies have leveragedLBSNs for understanding cultural differences betweenthe behaviors of users from different countries.

In the past decade, Yelp has become a worldwideonline business review site, which records millions ofreviews and business preferences of users from differentcountries. With the Yelp Open Dataset, numerousresearchers have conducted studies on Yelp. Byers et

al. [6] studied the correlation between the Grouponbehavior of a business and the user rating distributionto this business. On the Yelp Open Dataset, fakereviews or malicious reviewers were filtered out byanalyzing the texts of reviews [36–38]. Moreover,[17, 39] leveraged the rating and category or businesspreference extracted from the keywords or sentencesin the text to analyze users’ appetite and understandusers’ feedback. Furthermore, text and rating were alsoused together to understand users’ feedback [40]. Inthis work, we combine the text, rating and reviews onYelp to form a comprehensive understanding of userbehaviors from two cultural clusters.

To understand user behavior from a cross-culturalperspective, a preliminary work [41] reported thatdifferent cultural backgrounds has impacts at theindividual level on IT acceptance. To our knowledge,[5] is the first work that applied cultural analysismethods while studying the self-disclosure behavior onOSNs. They conducted a survey on Facebook andexplored the differences in individual willingness toself-disclosure between American and German users.[42] also analyzed the daily habits of school pupilsin Germany and China. Unfortunately, both studies[5, 42] relied on data collected from online surveys,which were quite limited. Garcia-Gavilanes et al.[43] considered the text-based “big data” on Twitter toascertain whether a strong relationship exists betweenusers’ behavior on Twitter and traditional culturaltheories. Even though [43] explored a large scale ofusers, aforementioned works [5, 42, 43] lacked richfeatures in their data, which are largely based oneither answers to a questionnaire or the “free text” onTwitter. In addition, for [43], they cared about therelationship between users’ online behavior and realculture phenomenon but did little about the differencesin users’ behavior between certain culture clusters, (i.e.,cross-cultural behavior). To narrow the gap betweencomprehensive data and online cross-cultural behavior,our work combines social connectivity, spatiotemporalpatterns, and writing styles of users on Yelp tounderstand the differences and similarities between


American and German users. We build on these pastworks to study the feasibility of using review-basedwebsites to analyze cultural differences between certaincultural clusters.

6 Conclusion

By referring to the Yelp Open Dataset, we use thebehavioral information of massive users to explorethe differences between American and German users.We find that the major differences are categorypreference, mobility pattern, friendship distribution,rating preference and writing style. We utilize theresults to extract several human behavior patterns andgeneralize their features. In addition, we build aclassification model to detect where a certain usercomes from based on the extracted features. With thismodel, we validate our analysis results and gain a betterunderstanding of the importance of various feature setsin forming a human behavior pattern on Yelp.

The cultural causes of user online behavior shouldbe studied in future work. Expanding the variety ofsocial platforms, for example, to other LBSNs likeFoursquare, would be an important study for us toanalyze the cultural differences, in combination with thepresent study using Yelp. We plan to use cross-OSNlinks [44] to obtain a user’s activity data from differentOSN sites. Meanwhile, an offline user study is also inour plans. We aim to build an overall behavior patternof cultural consequences which can be applied to peoplewith different cultural backgrounds.

Acknowledgment

This work is sponsored by National Natural ScienceFoundation of China (No. 61602122, No. 71731004),Natural Science Foundation of Shanghai (No.16ZR1402200), Shanghai Pujiang Program (No.16PJ1400700), EU FP7 IRSES MobileCloud project(No. 612212) and Lindemann Foundation (No. 12-2016).

References

[1] Harry C Triandis, Robert Bontempo, Marcelo J Villareal,Masaaki Asai, and Nydia Lucca. Individualismand collectivism: Cross-cultural perspectives on self-ingroup relationships. Journal of personality and SocialPsychology, 54(2):323, 1988.

[2] Rebecca Gumbrell-McCormick and Richard Hyman.Embedded collectivism? Workplace representation inFrance and Germany. Industrial Relations Journal,37(5):473–491, 2006.

[3] Matthias Ehrgott, Felix Reimann, Lutz Kaufmann, andCraig R Carter. Social sustainability in selecting emergingeconomy suppliers. Journal of business ethics, 98(1):99–119, 2011.

[4] Long Jin, Yang Chen, Tianyi Wang, Pan Hui, andAthanasios V. Vasilakos. Understanding User Behavior inOnline Social Networks: A Survey. IEEE CommunicationsMagazine, 51(9):144–150, 2013.

[5] Hanna Krasnova and Natasha F Veltri. Privacy calculuson social networking sites: Explorative evidence fromGermany and USA. In Proc. of the 43rd HawaiiInternational Conference on System Sciences (HICSS),pages 1–10. IEEE, 2010.

[6] John W Byers, Michael Mitzenmacher, and GeorgiosZervas. The groupon effect on yelp ratings: a root causeanalysis. In Proc. of the 13th ACM conference on ElectronicCommerce, pages 248–265, 2012.

[7] Marisa Affonso Vasconcelos, Saulo Ricci, Jussara Almeida,Fabrıcio Benevenuto, and Virgılio Almeida. Tips, donesand todos: uncovering user profiles in foursquare. In Proc.of the fifth ACM WSDM, pages 653–662, 2012.

[8] Yang Chen, Yuxi Yang, Jiyao Hu, and Chenfan Zhuang.Measurement and Analysis of Tips in Foursquare. In Proc.of IEEE PerCom Workshops, 2016.

[9] Y. Chen, J. Hu, H. Zhao, Y. Xiao, and P. Hui. Measurementand Analysis of the Swarm Social Network With Tens ofMillions of Nodes. IEEE Access, 6:4547–4559, 2018.

[10] Terence Chen, Mohamed Ali Kaafar, and Roksana Boreli.The Where and When of Finding New Friends: Analysisof a Location-based Social Discovery Network. In Proc. ofAAAI ICWSM, 2013.

[11] Rong Xie, Yang Chen, Shihan Lin, Tianyong Zhang,Yu Xiao, and Xin Wang. Understanding Skout Users’Mobility Patterns on a Global Scale: A Data-Driven Study.World Wide Web Journal, 2018.

[12] Yue Huang, Yang Chen, Qian Zhou, Jin Zhao, and XinWang. Where Are We Visiting? Measurement and Analysisof Venues in Dianping. In Proc. of IEEE InternationalConference on Communications (ICC), 2016.

[13] Qingyuan Gong, Yang Chen, Xinlei He, Zhou Zhuang,Tianyi Wang, Hong Huang, Xin Wang, and XiaomingFu. DeepScan: Exploiting Deep Learning for MaliciousAccount Detection in Location-Based Social Networks.IEEE Communications Magazine, 2018.

[14] Seymour Martin Lipset. Some social requisites ofdemocracy: Economic development and politicallegitimacy. American political science review, 53(1):69–105, 1959.


[15] Elke U Weber, Christopher K Hsee, and JoannaSokolowska. What folklore tells us about risk and risktaking: Cross-cultural comparisons of American, German,and Chinese proverbs. Organizational Behavior andHuman Decision Processes, 75(2):170–186, 1998.

[16] Michael Clyne. Cultural differences in the organizationof academic texts: English and German. Journal ofpragmatics, 11(2):211–241, 1987.

[17] Amy Hicks, Jeannie Horovitz, Madeline Hovarter, MayaMiki, Jennifer L Bevan, et al. Why people use Yelp. com:An exploration of uses and gratifications. Computers inHuman Behavior, 28(6):2274 – 2279, 2012.

[18] Jure Leskovec and Rok Sosic. SNAP: A General-PurposeNetwork Analysis and Graph-Mining Library. ACMTransactions on Intelligent Systems and Technology (TIST),8(1):1, 2016.

[19] Lawrence Page, Sergey Brin, Rajeev Motwani, and TerryWinograd. The PageRank citation ranking: Bringing orderto the web. Technical report, Stanford InfoLab, 1999.

[20] Xiaohan Zhao, Alessandra Sala, Christo Wilson, XiaoWang, Sabrina Gaito, Haitao Zheng, and Ben Y Zhao.Multi-scale dynamics in a massive online social network.In Proc. of ACM IMC, pages 171–184, 2012.

[21] Yong-Yeol Ahn, Seungyeop Han, Haewoon Kwak, SueMoon, and Hawoong Jeong. Analysis of topologicalcharacteristics of huge online social networking services.In Proc. of WWW, pages 835–844, 2007.

[22] JW Pennebaker, RJ Booth, RL Boyd, and ME Francis.Linguistic inquiry and word count: LIWC2015. Austin,TX: Pennebaker Conglomerates, 2015.

[23] Markus Wolf, Andrea B Horn, Matthias R Mehl,Severin Haug, James W Pennebaker, and Hans Kordy.Computergestutzte quantitative Textanalyse: Aquivalenzund Robustheit der deutschen Version des LinguisticInquiry and Word Count. Diagnostica, 54(2):85–98, 2008.

[24] James W Pennebaker and Martha E Francis. Cognitive,Emotional, and Language Processes in Disclosure.Cognition & Emotion, 10(6):601–626, 1996.

[25] K David Roach and Paul R Byrne. A cross-culturalcomparison of instructor communication in American andGerman classrooms. Communication Education, 50(1):1–14, 2001.

[26] Eunjoon Cho, Seth A Myers, and Jure Leskovec.Friendship and mobility: user movement in location-basedsocial networks. In Proc. of ACM KDD, pages 1082–1090,2011.

[27] Rong Xie, Yang Chen, Qinge Xie, Yu Xiao, and XinWang. We Know Your Preferences in New Cities:Mining and Modeling the Behavior of Travelers. IEEECommunications Magazine, 2018.

[28] Tianqi Chen and Carlos Guestrin. XGBoost: A ScalableTree Boosting System. In Proc. of ACM KDD, pages 785–794, 2016.

[29] Mark Hall, Eibe Frank, Geoffrey Holmes, BernhardPfahringer, Peter Reutemann, and Ian H. Witten. TheWEKA Data Mining Software: An Update. SIGKDDExplor. Newsl., 11(1):10–18, 2009.

[30] Tom Fawcett. An introduction to ROC analysis. PatternRecognition Letters, 27(8):861–874, 2006.

[31] Quinn McNemar. Note on the sampling error of thedifference between correlated proportions or percentages.Psychometrika, 12(2):153–157, 1947.

[32] Yiming Yang and Jan O. Pedersen. A Comparative Studyon Feature Selection in Text Categorization. In Proc. ofthe 14th International Conference on Machine Learning(ICML), pages 412–420, 1997.

[33] Giorgio Topa. Social interactions, local spilloversand unemployment. The Review of Economic Studies,68(2):261–295, 2001.

[34] Anne LJ Ter Wal and Ron A Boschma. Applying socialnetwork analysis in economic geography: framing somekey analytic issues. The Annals of Regional Science,43(3):739–756, 2009.

[35] David Liben-Nowell, Jasmine Novak, Ravi Kumar,Prabhakar Raghavan, and Andrew Tomkins. GeographicRouting in Social Networks. Proc. of the NationalAcademy of Sciences of the United States of America,102(33):11623–11628, 2005.

[36] Arjun Mukherjee, Vivek Venkataraman, Bing Liu, andNatalie S Glance. What yelp fake review filter might bedoing? In Proc. of AAAI ICWSM, 2013.

[37] Michael Luca and Georgios Zervas. Fake It Till YouMake It: Reputation, Competition, and Yelp Review Fraud.Management Science, 62(12):3412–3427, 2016.

[38] Yuanshun Yao, Bimal Viswanath, Jenna Cryan, HaitaoZheng, and Ben Y Zhao. Automated Crowdturfing Attacksand Defenses in Online Review Systems. In Proc. of ACMCCS, pages 1143–1158, 2017.

[39] Weranuj Ariyasriwatana and Luz Marina Quiroga.A thousand ways to say ‘Delicious!’–Categorizingexpressions of deliciousness from restaurant reviews on thesocial network site Yelp. Appetite, 104:18 – 32, 2016.

[40] Julian McAuley and Jure Leskovec. Hidden Factors andHidden Topics: Understanding Rating Dimensions withReview Text. In Proc. of the 7th ACM Conference onRecommender Systems, pages 165–172, 2013.

[41] Mark Srite and Elena Karahanna. The role of espousednational cultural values in technology acceptance. MISquarterly, pages 679–704, 2006.


[42] Xiaoming Fu, Hong Huang, Xiang-Yang Li, Haisheng Tan,and Jie Tang. A Comparative Analysis of School Pupils’Daily Habits in Germany and China. In Proc. of IEEEINFOCOM Workshops, 2018.

[43] Ruth Garcia-Gavilanes, Daniele Quercia, and AlejandroJaimes. Cultural dimensions in twitter: Time, individualism

and power. In Proc. of AAAI ICWSM, 2013.

[44] Qingyuan Gong, Yang Chen, Jiyao Hu, Qiang Cao, PanHui, and Xin Wang. Understanding Cross-site Linking inOnline Social Networks. ACM Transactions on the Web,2018.

Chenxi Yang is an undergraduate studentwithin the School of Computer Science atFudan University, China. She has been aresearch assistant in the Mobile Systemsand Networking (MSN) group since 2016.She visited Peking University (in 2017) asa research intern. Her research interestsinclude massive data analytics, machine

learning and edge computing.

Yang Chen is an Associate Professorwithin the School of Computer Science atFudan University, China. He leads theMobile Systems and Networking (MSN)group since 2014. Before joining Fudan,he was a postdoctoral associate at theDepartment of Computer Science, DukeUniversity, USA, where he served as

Senior Personnel in the NSF MobilityFirst project. FromSeptember 2009 to April 2011, he has been a research associateand the deputy head of Computer Networks Group, Instituteof Computer Science, University of Gottingen, Germany. Hereceived his B.S. and Ph.D. degrees from Department ofElectronic Engineering, Tsinghua University in 2004 and 2009,respectively. He visited Stanford University (in 2007) andMicrosoft Research Asia (2006-2008) as a visiting student.His research interests include online social networks, Internetarchitecture and mobile computing. He is serving as anEditorial Board Member of the Transactions on EmergingTelecommunications Technologies (ETT) and IEEE Access. Heis a senior member of the IEEE.

Qingyuan Gong received her B.S. degreein Computer Science from ShandongNormal University in 2012. She is nowa PhD candidate in Computer Science atFudan University, China. Her researchinterests include social network analytics,network security, and machine learning.She published referred papers in ACM

Transactions on the Web, IEEE Communications Magazine andIEEE ICPP. She has been a visiting student at the University ofGottingen (2015) and Tsinghua University (2016).

Xinlei He received his B.S. degree inComputer Science from Fudan Universityin 2017. He is now a master studentin Computer Science at Fudan University.He has been a research assistant in theMobile Systems and Networking (MSN)group since 2015. His research interestsinclude social computing and data mining.

Yu Xiao received her doctoral degree(with distinction) in computer sciencefrom Aalto University in January 2012.Before that, she got her Master andBachelor degrees in computer scienceand technology from Beijing Universityof Posts and Telecommunications, China.She is currently an assistant professor

in Department of Communications and Networking, AaltoUniversity where she leads the mobile cloud computinggroup. Her research interests include edge computing, mobilecrowdsensing, and energy-efficient wireless networking. Herwork has received 3 best paper awards from IEEE/ACMconferences. She is also a recipient of the 3-year postdoc grantfrom Academy of Finland.

Yuhuan Huang is an assistant professorwithin the Faculty of European Languagesand Cultures at Guangdong University ofForeign Studies. She received her doctoraldegree (with distinction) in InterculturalGerman Studies from the University ofGottingen, Germany in May 2017. Beforethat, she received her Master and Bachelor

degrees in German Studies from Beijing Foreign StudiesUniversity, China. Her research interests include Germanlanguage and culture as well as intercultural communication.

Xiaoming Fu is a full professor ofcomputer science at the University ofGottingen. He received his Ph.D. fromTsinghua University in 2000. He wasthen a research staff at TU Berlin beforemoving to Gottingen in 2002, where hehas been a professor and head of computernetworks group since 2007. His research

interests lie in networked systems and applications, includingmobile and cloud computing, social networks and big data


analysis. He currently serves on the editorial boards of IEEECommunications Magazine, IEEE Transactions on Network and

Service Management, and Computer Communications.

understanding the behavioral differences between american...

Documents