Transcription

Online Actions with Offline Impact: How Online SocialNetworks Influence Online and Offline User BehaviorTim AlthoffPranav JindalJure LeskovecStanford UniversityStanford UniversityStanford [email protected] of today’s most widely used computing applications utilizesocial networking features and allow users to connect, follow eachother, share content, and comment on others’ posts. However, despite the widespread adoption of these features, there is little understanding of the consequences that social networking has on userretention, engagement, and online as well as offline behavior.Here, we study how social networks influence user behavior ina physical activity tracking application. We analyze 791 milliononline and offline actions of 6 million users over the course of 5years, and show that social networking leads to a significant increase in users’ online as well as offline activities. Specifically,we establish a causal effect of how social networks influence userbehavior. We show that the creation of new social connections increases user online in-application activity by 30%, user retentionby 17%, and user offline real-world physical activity by 7% (about400 steps per day). By exploiting a natural experiment we distinguish the effect of social influence of new social connections fromthe simultaneous increase in user’s motivation to use the app andtake more steps. We show that social influence accounts for 55%of the observed changes in user behavior, while the remaining 45%can be explained by the user’s increased motivation to use the app.Further, we show that subsequent, individual edge formations in thesocial network lead to significant increases in daily steps. These effects diminish with each additional edge and vary based on edge attributes and user demographics. Finally, we utilize these insights todevelop a model that accurately predicts which users will be mostinfluenced by the creation of new social network connections.1.INTRODUCTIONSocial network features are central to many of today’s computing applications. Many successful websites and apps use social networking features to appeal to their users, allowing them to interact,form social connections, post updates, spread content, and comment on other’s posts. Social networking features are ubiquitousand are not only used by online social networks, such as Facebookand Twitter. For example, news reading, online education, music listening, book reading, diet and weight loss, physical activityPermission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others thanACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permissionand/or a fee. Request permissions from [email protected] 2017, February 06-10, 2017, Cambridge, United Kingdomc 2017 ACM. ISBN 978-1-4503-4675-7/17/02. . . tracking, and many other types of modern computing applicationsall heavily rely on social networking.Recent research has made great advancements towards understanding of fundamental structural properties [36, 41], growth [35],navigability [31, 37], community structure [11, 50], informationdiffusion [12, 19], influence maximization [30], social capital [17,27], and social influence [47] in online social networks. However,the impact of the online social networks on user behavior remainselusive. For example, little is known about whether and to what degree online social networking features influence user engagement,increase user retention, and change behavior within the immediateapplication as well as in the real-world. Furthermore, it is not clearwhether social networking features simply attract users that wouldbe more active and more engaged even if these features were absent, and whether social networks actually influence user online aswell as offline behavior.Existing studies of social influence in social networks have mainlybeen restricted to measuring online behaviors and outcomes such asthe adoption of apps [8, 9], downloads of content [47], voting oncontent [42], and resharing of content [43, 44]. However, manyimportant behaviors and outcomes pertain to the offline world including political mobilization [16], physical activity [5, 6, 25, 34],food intake [24], mental health [1, 23], obesity [20], and smoking [21]. In order to study offline behaviors, researchers have usedproxies that are observable online: for example, posting in a particular forum or an app to measure dieting choices [24, 40], suicidal thoughts [23], helping behavior [2] and charitable behavior [4].However, one has to trust that the self-reports observed online correspond to objective behaviors and many studies have shown largebiases of such self-reports [14, 29, 49].Estimating the influence of social networks on online as well asoffline behavior is challenging due to unobserved counterfactualbehavior, where one cannot observe a user’s behavior had they notjoined the social network. Furthermore, selection effects complicate causal estimation from observational data [7, 8, 32]. For example, social network users could exhibit different behaviors due to(1) a selection effect of what kind of users would choose to join thesocial network, or (2) an actual influence effect of the social network on their behavior. Often the mere act of being part of a socialnetwork already means that these users are particularly motivatedto use the app and take more steps. In many contexts all of theseeffects are acting simultaneously (e.g., [13] for health behaviors),which creates further challenges and makes causal identification ofeffects even harder.Present work. In this paper, we study the influence of social networks on users’ online as well as offline behavior. We study userbehavior in a smartphone physical activity tracking application,that allows us to observe users’ in-application engagement as well

as the offline real-world physical activity measured through smartphone accelerometry. Specifically, we use data from the AzumioArgus app, which tracks exercise and physical activity of 6 millionusers worldwide over the course of 5 years (2011-2016). Duringthis time, users created 631 million activity posts (e.g., runs, sleep,cycling, yoga, etc.) by actively opening the app and self-reportingthe activity. In contrast, physical activity is passively collectedthrough smartphone sensors in the form of step measurements without the need for self-reports and active user engagement. Our dataset additionally contains 160 million days of passive steps trackingadding up to 824 billion total steps taken. Therefore, we distinguishbetween activity posts as a measure of online app engagement andsteps taken as a measure of offline physical activity behavior. To thebest of our knowledge this is the largest dataset on human activitytracking and social network interactions to date (e.g., ten thousandtimes more users and a million times more activity tracking thancomparable studies [26]).An internal social network was introduced in the application inNovember 2013 and since then a subset of users have chosen to joinand engage in the social network. Using the data, we quantify thecausal effect of the social network on user behavior by harnessinga novel natural experiment on delayed social network edge formation. In particular, we distinguish the causal effect of social influence of a new network connection from the simultaneous increasein motivation of the user to use the app (i.e., a selection effect). Weshow that social influence does explain 55% of the observed average effect, while the remaining 45% of the observed effects are dueto the increased motivation.We find that joining the social network has significant positiveeffect on online and offline user behavior that diminish over time.Users of the social network are 30% more active in the app, 17%less likely to drop out of the app within one year, and 7% morephysically active compared to a matched control group, and weshow that these effects last over long periods of several months.Further, we estimate the effect of subsequent, individual edge formations in the social network. We observe temporary increasesin offline physical activity that diminish with each additional connection and are larger for friend connections than follower connections. Further, these average increases are larger for the initiatorof the connection than its recipient, and the effect varies with age,gender, weight, and prior physical activity level. Finally, we utilize these insights to develop a model to predict which users will bemost influenced by the creation of new social network connectionsand show that the proposed factors explain a significant fractionof the variability in user’s behavior change. We conclude by discussing related work on online social networks and social media inthe context of health applications.In summary, the main contributions of this work include:1. We study the causal impact of social network features on userbehavior using the largest activity tracking dataset to date.2. We show how online social networks shape online as well asoffline user behavior, including user engagement, retention,and real-world physical activity.3. We employ natural experiments, difference-in-difference models, and matching-based observational studies to disentangleselection effects from causal social network effects.While we focus on physical activity and health behaviors in thiswork, our methods are more generally applicable to other offlineand online activities.2.DATASET DESCRIPTIONWe use a dataset of 6 million individuals from over 100 different countries using the Argus smartphone app by Azumio whichallows users to track their daily activities. Over a time period of 5years we observe 631 million self-reported activity posts (includingrunning, walking, sleep, heart rate, yoga, cycling, weight, etc.) and160 million days of steps tracking (objectively measured throughthe smartphone accelerometers) between January 2011 and January2016. Table 1 further summarizes the worldwide dataset and showsthat the distribution of age, gender, and weight is fairly representative of the overall population in many developed countries. Forexample, the median age in our dataset for U.S. users is 34 yearswhich is close to the official estimate of 37 years. Furthermore,28% of these users are obese closely matching previous publishedestimates of 30-38%. We also include plots of the degree distribution and distribution of edge inter-creation times in the onlineappendix [3].Throughout the paper, we distinguish between the number ofaccelerometer-defined steps (physical activity; offline) and the number of posts the user creates within the app each corresponding to aself-reported action such as running, cycling or sleeping (in-app activity; online). We use “activity” to refer to offline physical activityand “posts” to refer to online in-application user activity.In November of 2013, a social network feature was introducedin the app allowing both bi-directional friend connections (after approval of a friend request by the receiver) as well as uni-directionalfollower connections (without need for approval). New social connections result in receiving notifications of the other person’s activity posts (e.g., runs and walks). Furthermore, the app includes atimeline-like activity feed that then includes the activity of the newfriend and enables the user to comment on others’ activity postsfor encouragement and support. During our observation period, alledges in the network were created organically without any influence of friend recommendation algorithms.This dataset uniquely enables the study of how social networkfeatures impact user behavior. It captures both online (actions withinthe app) as well as offline user behaviors (physical steps taken in theoffline world). Further, it has two relevant properties: (1) The social network was introduced after two years of observing behaviorwithout any social interactions. (2) The delays with which friendship requests get accepted form a natural experiment, which allowsus to disentangle influence of social networking features from simple selection effects. Our analyses exploit these properties to carryout two natural experiments that provide novel insights into howuser behavior is shaped by social network interactions. Lastly, thelarge-scale nature of our dataset—about two million times moreposts than previously published research [26]—allows us to studyvarious kinds of heterogeneous effects, for example across age,gender, BMI, previous activity level. Data handling and analysiswas conducted in accordance with the guidelines of the appropriateInstitutional Review Board.3.DISTINGUISHING INTRINSIC MOTIVATION FROM SOCIAL INFLUENCEIn this section, we quantify the average influence that a singlesocial network edge has on the person who sends the friendshiprequest. We present a novel approach based on a natural experimentto distinguish the effect of (1) increased user’s motivation whenadding a new edge, from (2) establishing a social network edgethat influences the user to change their behavior.Challenges of distinguishing motivation from influence. To illustrate the challenge of distinguishing the two effects, Figure 1

Observation periodIntroduction of the social network# total users# total online and offline activities# activity posts (online engagement)# users tracking steps# days of steps tracking (offline activity)# total steps tracked# users in the social network# edges in the social networkMedian age% users female% underwe