The Role of Social Networks in Information Diffusion

Eytan Bakshy
Facebook 1601 Willow Rd. Menlo Park, CA 94025

Itamar Rosenn
Facebook 1601 Willow Rd. Menlo Park, CA 94025

ebakshy@fb.com Cameron Marlow

itamar@fb.com Lada Adamic

Facebook 1601 Willow Rd. Menlo Park, CA 94025

cameron@fb.com

ladamic@umich.edu
38]. Despite the wide availability of data from online social networks, identifying inuence remains a challenge. Individuals tend to engage in similar activities as their peers, so it is often impossible to determine from observational data whether a correlation between two individuals behaviors exists because they are similar or because one persons behavior has inuenced the other [5, 32, 39]. In the context of information diusion, two people may disseminate the same information as each other because they possess the same information sources, such as web sites or television, that they consume regularly [3, 38]. Moreover, homophily the tendency of individuals with similar characteristics to associate with one another [1, 28, 34] creates diculties for measuring the relative role of strong and weak ties in information diusion, since people are more similar to those with whom they interact often [22, 34]. On one hand, pairs of individuals who interact more often have greater opportunity to inuence one another and have more aligned interests, increasing the chances of contagion [11, 27]. However, this commonality amplies the potential for confounds: those who interact more often are more likely to have increasingly similar information sources. As a result, inferences made from observational data may overstate the importance of strong ties in information spread. Conversely, individuals who interact infrequently have more diverse social networks that provide access to novel information [12, 22]. But because contact between such ties is intermittent, and the individuals tend to be dissimilar, any particular piece of information is less likely to ow across weak ties [14, 37]. Historical attempts to collect data on how often pairs of individuals communicate and where they get their information have been prone to biases [10, 33], further obscuring the empirical relationship between tie strength and diusion. Confounding factors related to homophily can be addressed using controlled experiments, but experimental work has thus far been conned to the spread of highly specic information within limited populations [6, 13]. In order to understand how information spreads in a real-world environment, we wish to examine a setting where a large population of individuals frequently exchange information with their peers. Facebook is the most widely used social networking service in the world, with over 800 million people using the service each month. For example, in the United States, 54% of adult Internet users are on Facebook [26].

University of Michigan 105 S. State St. Ann Arbor, MI 48104

ABSTRACT
Online social networking technologies enable individuals to simultaneously share information with any number of peers. Quantifying the causal eect of these mediums on the dissemination of information requires not only identication of who inuences whom, but also of whether individuals would still propagate information in the absence of social signals about that information. We examine the role of social networks in online information diusion with a large-scale eld experiment that randomizes exposure to signals about friends information sharing among 253 million subjects in situ. Those who are exposed are signicantly more likely to spread information, and do so sooner than those who are not exposed. We further examine the relative role of strong and weak ties in information propagation. We show that, although stronger ties are individually more inuential, it is the more abundant weak ties who are responsible for the propagation of novel information. This suggests that weak ties may play a more dominant role in the dissemination of information online than currently believed.

Categories and Subject Descriptors

H.1.2 [Models and Principles]: User/Machine Systems; J.4 [Social and Behavioral Sciences]: Sociology

General Terms
Experimentation, Measurement, Human Factors

Keywords
social inuence, tie strength, causality

1.

INTRODUCTION

Social inuence can play a crucial role in a range of behavioral phenomena, from the dissemination of information, to the adoption of political opinions and technologies [23, 42], which are increasingly mediated through online systems [17, Part of this research was performed while the author was a student at the University of Michigan.
Copyright is held by the International World Wide Web Conference Committee (IW3C2). Distribution of these papers is limited to classroom use, and personal use by others. WWW 2012, April 1620, 2012, Lyon, France. ACM 978-1-4503-1229-5/12/04.

Those American users on average maintain 48% of their real world contacts on the site [26], and many of these individuals regularly exchange news items with their contacts [38]. In addition, interaction among users is well correlated with selfreported intimacy [18]. Thus, Facebook represents a broad online population of individuals whose online personal networks reect their real-world connections, making it an ideal environment to study information contagion. We use an experimental approach on Facebook to measure the spread of information sharing behaviors. The experiment randomizes whether individuals are exposed via Facebook to information about their friends sharing behavior, thereby devising two worlds under which information spreads: one in which certain information can only be acquired external to Facebook, and another in which information can be acquired within or external to Facebook. By comparing the behavior of individuals within these two conditions, we can determine the causal eect of the medium on information sharing. The remainder of this paper is organized as follows. We further motivate our study with additional related work in Section 2. Our experimental design is described in Section 3. Then, in Section 4 we discuss the causal eect of exposure to content on the newsfeed, and how friends sharing behavior is correlated in time, irrespective of social inuence via the newsfeed. Furthermore, we show that multiple sharing friends are predictive of sharing behavior regardless of exposure on the feed, and that additional friends do indeed have an increasing causal eect on the propensity to share. In Section 5 we discuss how tie strength relates to inuence and information diusion. We show that users are more likely to have the same information sources as their close friends, and that simultaneously, these close friends are more likely to inuence subjects. Using the empirical distribution of tie strength in the network, we go on to compute the overall eect of strong and weak ties on the spread of information in the network. Finally, we discuss the implications of our work in Section 6.

contagion event is a friend, such data does not tell us about the relative importance of social networks in information diffusion. For example, consider the spread of news. In Bradley Greenbergs classsic study of media contagion [24], 50% of respondents learned about the Kennedy assassination via interpersonal ties. Despite the substantial word-of-mouth spread, it is clear that all of the respondents would have gotten the news at a slightly later point in time (perhaps from the very same media outlets as their contacts), had they not communicated with their peers. Therefore, a complete understanding of the importance of social networks in information diusion not only requires us to identify sources of interpersonal contagion, but also requires a counterfactual understanding of what would happen if certain interactions did not take place.
Observable
Facebook Feed

Unobservable
External Correlation
Regular visitation to web sites

Story 1 Story 2 ... Story linking to page X ...

E-Mail

...

Instant Messaging

User visits page X

User shares page X on Facebook

2.

RELATED WORK

Online networks are focused on sharing information, and as such, have been studied extensively in the context of information diusion. Diusion and inuence have been modeled in blogs [2, 20, 25], email [31], and sites such as Twitter, Digg, and Flickr [8, 21, 29]. One particularly salient characteristic of diusion behavior is the correlation between the number of friends engaging in a behavior and the probability of adopting the behavior. This relationship has been observed in many online contexts, from the joining of LiveJournal groups [7], to the bookmarking of photos [15], and the adoption of user-created content [9]. However, as Anagnostopoulos, et al. [4] point out, individuals may be more likely to exhibit the same behavior as their friends because of homophily rather than as a result of peer inuence. Statistical techniques such as permutation tests and matched sampling [5] help control for confounds, but ultimately cannot resolve this fundamental problem [39]. Not all diusion studies must infer whether one individual inuenced another. For example, Leskovec et al. [30] study the explicit graph of product recommendations, Sun et al. [41] study cascading in page fanning, and Bakshy et al. [9] examine the exchange of user-created content. However, in all these studies, even if the source of a particular

Figure 1: Causal relationships that explain diusion-like phenomena. Information presented in users news feeds and other sharing behavior on facebook.com are observed. External events that cause users to be exposed to information outside of Facebook cannot be observed and may explain their sharing behavior. Our experiment blocks the causal relationship (dashed arrow) between the Facebook newsfeed and user visitation by randomly removing stories about friends sharing behavior in subjects feeds. Thus, our experiment allows us to compare situations where both inuence via the feed and external correlations exist (the feed condition), to situations in which only external correlations exist (the no feed condition).

3.

EXPERIMENTAL DESIGN AND DATA

Facebook users primarily interact with information through an aggregated history of their friends recent activity (stories), called the News Feed, or simply feed for short. Some of these stories contain links to content on the Web, uniquely identied by URLs. Our experiment evaluates how much exposure to a URL on the feed increases an individuals propensity to share that URL, beyond correlations that one might expect among Facebook friends. For example, friends with whom a user interacts more often may be more likely to visit sites that the user also visits. As a result, those friends may be more likely to share the same URL as the

(a)

(b)

Figure 2: An example of the Facebook News Feed interface for a hypothetical subject who has a link (highlighted in red) assigned to the (a) feed or (b) no feed condition.

user before she has the opportunity to share that content herself. Additional unobserved correlations may arise due to external inuence via e-mail, instant messaging, and other social networking sites. These causal relationships are illustrated in Figure 1. From the gure, one can see that all unobservable correlations can be identied by blocking the causal relationship between the Facebook feed and sharing. Our experiment therefore randomizes subjects with respect to whether they receive social signals about friends sharing behavior of certain Web pages via the Facebook feed.

3.1

Assignment Procedure

Subject-URL pairs are randomly assigned at the time of display to either the no feed or the feed condition. Stories that contain links to a URL assigned to the no feed condition for the subject are never displayed in the subjects feed. Those assigned to the feed condition are not removed from the feed, and appear in the subjects feed as normal (Figure 2). Pairs are deterministically assigned to a condition at the time of display, so any subsequent share of the same URL by any of a subjects friends is also assigned to the same condition. To improve the statistical power of our results, twice as many pairs were assigned to the no feed condition. Because removal from the feed occurs on a subject-URL basis, and we include only a small fraction of subject-URL pairs in the no feed condition, a shared URL is on average delivered to over 99% of its potential targets.

All activity relating to subject-URL pairs assigned to either experimental condition is logged, including feed exposures, censored exposures, and clicks to the URL (from the feed or other sources, like messaging). Directed shares, such as a link that is included in a private Facebook message or explicitly posted on a friends wall, are not aected by the assignment procedure. If a subject-URL pair is assigned to an experimental condition, and the subject clicks on content containing that URL in any interface other than the feed, that subject-URL pair is removed from the experiment. Our experiment, which took place over the span of seven weeks, includes 253,238,367 subjects, 75,888,466 URLs, and 1,168,633,941 unique subject-URL pairs.

3.2

Ensuring Data Quality

Threats to data quality include using content that was or may have been previously seen by subjects on Facebook prior to the experiment, content that subjects may have seen through interfaces on Facebook other than feed, spam, and malicious content. We address these issues in a number of ways. First, we only consider content that was shared by the subjects friends only after the start of the experiment. This enables our experiment to accurately capture the rst time a subject is exposed to a link in the feed, and ensures that URLs in our experiment more accurately reect content that is primarily being shared contemporaneously with the timing of the experiment. We also exclude potential subject-

Demographic Feature (% of subjects) Gender Female Male Unspecified Age 17 or younger 18-25 26-35 36-45 46 or older Country (top 10 & other) United States Turkey Great Britain Italy France Canada Indonesia Philippines Germany Mexico 226 Others

feed

no feed

51.6% 46.7% 1.5% 12.8% 36.4% 27.2% 13.0% 10.6% 28.9% 6.1% 5.1% 4.2% 3.8% 3.7% 3.7% 2.1% 2.3% 2.0% 37.5%

51.4% 47.0% 1.5% 13.1% 36.1% 26.9% 12.9% 10.9% 29.1% 5.8% 5.2% 4.1% 3.9% 3.8% 3.5% 2.3% 2.3% 2.1% 37.7%

Table 1: Summary of demographic features of subjects assigned to the feed (N = 160, 688, 092) and no feed (N = 218, 743, 932) condition. Some subjects may appear in both columns.

URL pairs where the subject had previously clicked on the URL via any interface on the site at any time up to two months prior to exposure, or any interface other than the feed for content assigned to the no feed condition. Finally, we use the Facebooks site integrity system [40] to classify and remove URLs that may not reect ordinary users purposeful intentions of distributing content to their friends.

and tie strength, which are analyzed in Sections 4 and 5. Alternatively, the dierence in probabilities can be viewed as a ratio (the relative risk ratio), which quanties how many times more likely an individual is to share as a result of being exposed to content on the feed. Although the assignment is completely random, subjects and URLs may dier in ways that impact our measurements. For example, certain users may be highly active on Facebook, so that they are assigned to experimental conditions more often than other users. If these users were to vary signicantly in terms of their information sharing propensities, such as sharing or re-sharing greater or fewer links than others, the disproportionate inclusion of these users may bias our measurements and threaten the population validity of our ndings. Similarly, very popular URLs may also introduce biases; they may be more or less likely to be re-shared because of their inherent appeal or more likely to be discovered independently of Facebook because of their relative popularity amongst friends. To provide control for these biases, we use bootstrapped averages clustered by the subject or URL. We nd that in all of our analyses, clustering by the URL rather than the subject yields nearly identical probability estimates that have marginally wider condence intervals, so we have chosen to present our results using means and 95% condence intervals clustered by URL. Risk ratios are obtained using the 95% bootstrapped condence intervals of likelihood of sharing in the feed and no feed conditions. To compute the lower bound of the ratio, we divide the lower bound of the probability of sharing in the feed condition by the upper bound for the no feed condition. The upper bound of the ratio is computed by dividing the upper bound in the feed condition by the lower bound of the no feed condition. The additive analog of the same procedure is used to obtain condence intervals for probability dierences.

3.3

Population

4.

The experimental population consists of a random sample of all Facebook users who visited the site between August 14th to October 4th 2010, and had at least one friend sharing a link. At the time of the experiment, there were approximately 500 million Facebook users logging in at least once a month. Our sample consists of approximately 253 million of these users. All Facebook users report their age and gender, and a users country of residence can be inferred from the IP address with which she accesses the site. In our sample, the median and average age of subjects is 26 and 29.3, respectively. Subjects originate from 236 countries and territories, 44 of which have one million or more subjects. Additional summary statistics are given in Table 1, and show that subjects are assigned to the conditions in a balanced fashion.

HOW EXPOSURE TO SOCIAL SIGNALS AFFECTS DIFFUSION

We nd that subjects who are exposed to signals about friends sharing behavior are several times more likely to share that same information, and share sooner than those who are not exposed. To measure the relative increase in sharing due to exposure, we compute the risk ratio: the likelihood of sharing in the feed condition (0.191%) divided by the likelihood of sharing in the no feed condition (0.025%), and nd that individuals in the feed condition are 7.37 times more likely share (95% CI = [7.23, 7.72]). Although the probability of sharing upon exposure may appear small, it is important to note that individuals have hundreds of contacts online who may see their link, and that on average one out of every 12.5 URLs that are clicked on in the feed condition are subsequently re-shared.

3.4

Evaluating Outcomes

The assignment procedure allows us to directly compare the overall probability that subjects share links they were or were not exposed to on the feed. The causal eect of exposure via the Facebook feed on sharing is simply the expected probability of sharing in the feed condition minus the expected probability in the no feed condition. This quantity, known as the average treatment eect on the treated (or alternatively, the absolute risk increase), can vary when conditioning on other variables, including the number of friends

4.1

Temporal Clustering

Contemporaneous behavior among connected individuals is commonly used as evidence for social inuence processes (e.g. [4, 9, 8, 15, 16, 19, 20, 25, 29, 36, 43]). We nd that subjects who share the same link as their friends typically do so within a time that is proximate to their friends sharing time, even when no exposure occurs on Facebook. Figure 3 illustrates the cumulative distribution of information lags between the subject and their rst sharing friend, among

1.0 0.8

0.6 0.4 0.2 0.0 0 5 10 15 20 25 30

feed no feed

share time alter's share time (days)

tion. We rst match the share time of each URL in the feed condition with a share time of the URL in the no feed condition, sampling URLs in proportion to their relative abundances in the data. From this set of contrasts, we nd that the median sharing latency after a friend has already shared the content is 6 hours in the feed condition, compared to 20 hours when assigned to the no feed condition (Wilcoxon rank-sum test, p < 1016 ). The presence of strong temporal clustering in both experimental conditions illustrates the problem with inferring inuence processes from observations of temporally proximate behavior among connected individuals: regardless of access to social signals within a particular online medium, individuals can still acquire and share the same information as their friends, albeit at a slightly later point in time.

cumulative density

(a)
1.0 0.8

4.2

Effect of Multiple Sharing Friends

0.6 0.4 0.2 0.0 5 0 5 10 15 20 25 30

feed no feed

share time exposure time (days)

(b) Figure 3: Temporal clustering in sharing the same link as a friend in the feed and no feed conditions. (a) The dierence in sharing time between a subject and their rst sharing friend. (b) The dierence between the time at which a subject was rst to exposed (or was to be exposed) to the link and the time at which they shared. Vertical lines indicate one day and one week.

subjects who had shared a URL after their friends. The top panel shows the latency in sharing times between the subject and their friend for users in the feed and no feed condition. While a larger proportion of users in the feed condition share a link within the rst hour of their friends, the distribution of sharing times is strikingly similar. The bottom panel shows the dierences in time between when subjects shared and when they were (or would have been) rst exposed to their friends sharing behavior on the Facebook feed. The horizontal axis is negative when a subject had shared a link after a friend but had not yet seen that link on the feed. From this comparison, it is easy to see that users in the feed condition are most likely to share a link immediately upon exposure, while those who share it without seeing it in their feed will do so over a slightly longer period of time. To evaluate how exposure on the Facebook feed relates to the speed at which URLs appear to diuse, we consider URLs that were assigned to both the feed and no feed condi-

Classic models of social and biological contagion (e.g. [23, 35]) predict that the likelihood of infection increases with the number of infected contacts. Observational studies of online contagion [4, 9, 15, 30] not only nd evidence of temporal clustering, but also observe a similar relationship between the likelihood of contagion and the number of infected contacts. However, it is important to note that this correlation can have multiple causes that are unrelated to social inuence processes. For example, if a website is popular among friends, then a particularly interesting page is more likely to be shared by a users friends independent of one another. The positive relationship between the number of sharing friends and likelihood of sharing may therefore simply reect heterogeneity in the interestingness of the content, which is clustered along the network: the more popular a page is for a group of friends, the more likely it is that one would observe multiple friends sharing it. We rst show that, consistent with prior observational studies, the probability of sharing a link in the feed condition increases with the number of contacts who have already shared the link (solid line, Figure 4a). But the presence of a similar relationship in the no feed condition (grey line, Figure 4a) shows that an individual is more likely to exhibit the sharing behavior when multiple friends share, even if she does not necessarily observe her friends behavior. Therefore, when using observational data, the na conditional ve probability (which is equivalent to the probability of sharing in the feed condition) does not directly give the probability increase due to inuence via multiple sharing friends. Rather, such an estimate reects a mixture of internal inuence eects and external correlation. Our experiment allows us to directly measure the eect of the feed relative to external factors, computed as either the dierence or ratio between the probability of sharing in the feed and no feed conditions (Figure 4bc). While the dierence in sharing likelihood grows with the number of sharing friends, the relative risk ratio falls. This contrast suggests that social information in the feed is most likely to inuence a user to share a link that many of her friends have shared, but the relative impact of that inuence is highest for content that few friends are sharing. The decreasing relative eect is consistent with the hypothesis that having multiple sharing friends is associated with greater redundancy in information exposure, which may either be caused by homophily in visitation and sharing tendencies, or external inuence.

cumulative density

probability of sharing

0.030 0.025 0.020 0.015 0.010 0.005 0.000

condition feed no feed

0.030 0.025

10 8

p f eed p no feed

p f eed p no feed
1 2 3 4 5 6

0.020 0.015 0.010 0.005 0.000

6 4 2 0 1 2 3 4 5 6

number of sharing friends

number of sharing friends

number of sharing friends

(a)

(b)

(c)

Figure 4: Users with more friends sharing a Web link are themselves more likely to share. (a) The probability of sharing for subjects that were (feed) and were not (no feed) exposed to content increases as a function of the number sharing friends. (b) The causal eect of the feed is greater when subjects have more sharing friends (c) The multiplicative impact of the feed is greatest when few friends are sharing. Error bars represent the 95% bootstrapped condence intervals clustered on the URL.

5.

TIE STRENGTH AND INFLUENCE

cumulative fraction of ties in feed

Next, we examine the relationship between tie strength, inuence, and information diversity by combining the experimental data with users online and oine interactions. Following arguments originally proposed by Mark Granovetters seminal 1973 paper, The Strength of Weak Ties [22], empirical work linking tie strength and diusion often utilize the number of mutual contacts as proxies of interaction frequency. Rather than using the number of mutual contacts, which can be large for pairs of individuals who no longer communicate (e.g. former classmates), we directly measure the strength of tie between a subject and her friend in terms of four types of interactions: (i) the frequency of private online communication between the two users in the form of Facebook messages1 ; (ii) the frequency of public online interaction in the form of comments left by one user on another users posts; (iii) the number of real-world coincidences captured on Facebook in terms of both users being labeled by users as appearing in the same photograph; and (iv) the number of online coincidences in terms of both users responding to the same Facebook post with a comment. Frequencies are computed using data from the three months directly prior to the experiment. The distribution of tie strengths among subjects and their sharing friends can be seen in Figure 5.

1.00 0.98 0.96 0.94 0.92 0.90 0.88 0 10 20

type comments received messages received photo coincidences thread coincidences

30

40

50

tie strength
Figure 5: Tie strength distribution among friends displayed in subjects feeds using the four measurements. Points are plotted up to the 99.9th percentile. Note that the vertical axis is collapsed.

5.1

Effect of Tie Strength

sharing friend is a strong tie (Figure 6a). For example, subjects who were exposed to a link shared by a friend from whom the subject received three comments are 2.83 times more likely to share than subjects exposed to a link shared by a friend from whom they received no comments. For those who were not exposed, the same comparison shows that subjects are 3.84 times more likely to share a link that was previously shared by the stronger tie. The larger effect in the no feed condition suggests that tie strength is a stronger predictor of externally correlated activity than it is for inuence on feed. From Figure 6a, it is also clear that individuals are more likely to be inuenced by their stronger

We measure how the dierence in the likelihood of sharing a URL in the feed versus no feed conditions varies according to tie strength. To simplify our estimate of the eect of tie strength, we restrict our analysis to subjects with exactly one friend who had previously shared the link. In both conditions, a subject is more likely to share a link when her
1 We quantify message and comment interactions as the number of communication events the subject received from their friend. The number of messages and comments sent, and the geometric mean of communications sent and received, yielded qualitatively similar results, so we plot only the single directed measurement for the sake of clarity.

probability of sharing

probability of sharing

probability of sharing

condition feed no feed

0.006 0.004 0.002 0.000 0

0.006 0.004 0.002 0.000 0 1 2 3 4 5 6 7

0.006 0.004 0.002 0.000 0 1 2 3 4

probability of sharing

0.008

0.008

0.008

0.008 0.006 0.004 0.002 0.000 0 2 4 6 8 10 12 14

8 10 12

comments received

messages received

photo coincidences

thread coincidences

(a)
10 8 10 8 10 8 10 8

p f eed p no feed

p f eed p no feed

p f eed p no feed

6 4 2 0 0 2 4 6 8 10 12

6 4 2 0 0 1 2 3 4 5 6 7

6 4 2 0 0 1 2 3 4

p f eed p no feed photo coincidences

6 4 2 0 0 2 4 6 8 10 12 14

comments received

messages received

thread coincidences

(b) Figure 6: Strong ties are more inuential, and weak ties expose friends to information they would not have otherwise shared. (a) The increasing relationship between tie strength and the probability of sharing a link that a friend shared in the feed and no feed conditions. (b) The multiplicative eect of feed diminishes with tie strength, suggesting that exposure through strong ties may be redundant with external exposure, while weak ties carry information one might otherwise not have been exposed to. ties via the feed to share content that they would not have otherwise spread. Furthermore, our results extend Granovetters hypothesis that weak ties disseminate novel information into the context of media contagion. Figure 6b shows that the risk ratio of sharing between the feed and no feed conditions is highest for content shared by weak ties. This suggests that weak ties consume and transmit information that one is unlikely to be exposed to otherwise, thereby increasing the diversity of information propagated within the network. Under this categorization of strong and weak ties, the estimated total fraction of sharing events that can be attributed to weak and strong ties is the average treatment eect on the treated weighted by the proportion of URL exposures from each tie type: Tweak = ATET(0) f (0)
N

Tstrong =
i=1

ATET(i) f (i)

5.2

Collective Impact of Ties

Strong ties may be individually more inuential, but how much diusion occurs in aggregate through these ties depends on the underlying distribution of tie strength (i.e. Figure 5). Using the experimental data, we can estimate the amount of contagion on the feed generated by strong and weak ties. The causal eect of exposure to information shared by friends with tie strength k is given by the average treatment eect on the treated: ATET(k) = p(k, feed) p(k, no feed) To determine the collective impact of ties of strength k, we multiply this quantity by the fraction of links displayed in all users feeds posted by friends of tie strength k, denoted by f (k). In order to compare the impact of weak and strong ties, we must set a cuto value for the minimum amount of interaction required between two individuals in order to consider that tie strong. Setting the cuto at k = 1 (a single interaction) provides the most generous classication of strong ties while preserving some meaningful distinction between strong and weak ties, thereby giving the most inuence credit to strong ties.

We illustrate this comparison in Figure 7, and show that by a wide margin, the majority of inuence is generated by weak ties2 . Although we have shown that strong ties are individually more inuential, the eect of strong ties is not large enough to match the sheer abundance of weak ties.

6.

DISCUSSION

Social networks may inuence an individuals behavior, but they also reect the individuals own activities, interests, and opinions. These commonalities make it nearly impossible to determine from observational data whether any particular interaction, mode of communication, or social environment is responsible for the apparent spread of a behavior through a network. In the context of our study, there are three possible mechanisms that may explain diusion-like phenomena: (1) An individual shares a link on Facebook,
2 Note that for the purposes of this study, it is not necessary to model the eect of tie strength for users with multiple sharing friends, since stories of this kind only constitute 4.2% of links in the newsfeed, and their inclusion would not dramatically alter the balance of aggregate inuence by tie strength.

weak strong weak tie strength strong weak strong weak strong 0 20 40 60 80

% influence on feed

Figure 7: Weak ties are collectively more inuential than strong ties. Panels show the percentage of information spread by strong and weak ties for all four measurements of tie strength. Although the probability of inuence is signicantly higher for those that interact frequently, most contagion occurs along weak ties, which are more abundant.

and exposure to this information on the feed causes a friend to re-share that same link. (2) Friends visit the same web page and share a link to that web page on Facebook, independently of one another. (3) An individual shares a link within and external to Facebook, and exposure to the externally shared information causes a friend to share the link on Facebook. Our experiment determines the causal eect of the feed on the spread of sharing behaviors by comparing the likelihood of sharing under the feed condition (possible causes 1-3) with the likelihood under the no feed condition (possible causes 2-3). Our experiment generalizes Mark Granovetters predictions about the strength of weak ties [22] to the spread of everyday information. Weak ties are argued to have access to more diverse information because they are expected to have fewer mutual contacts; each individual has access to information that the other does not. For information that is almost exclusively embedded within few individuals, like job openings or future strategic plans, weak ties play a necessarily role in facilitating information ow. This reasoning, however, does not necessarily apply to the spread of widely available information, and the relationship between tie strength and information access is not immediately obvious. Our experiment sheds light on how tie strength relates to information access within a broader context, and suggests that weak ties, dened directly in terms of interaction propensities, diuse novel information that would not have otherwise spread. Although weak ties can serve a critical bridging function [22, 37], the inuence that weak ties exert has never before been measured empirically at a systemic level. We

nd that the majority of inuence results from exposure to individual weak ties, which indicates that most information diusion on Facebook is driven by simple contagion. This stands in contrast to prior studies of inuence on the adoption of products, behaviors or opinions, which center around the eect of having multiple or densely connected contacts who have adopted [6, 7, 14, 13]. Our results suggest that in large online environments, the low cost of disseminating information fosters diusion dynamics that are dierent from situations where adoption is subject to positive externalities or carries a high cost. Because we are unable to observe interactions that occur outside of Facebook, a limitation of our study is that we can only fully identify causal eects within the site. Correlated sharing in the no feed condition may occur because friends independently visit and share the same page as one another, or because one user is inuenced to share via an external communication channel. Although we are not able to directly evaluate the relative contribution of these two potential causes, our results allow us to obtain a bound on the eect on sharing behavior within the site. The probability of sharing in the no feed condition, which is a combination of similarity and external inuence, is an upper bound on how much sharing occurs because of homophily-related eects. Likewise, the dierence in the probability of sharing within the feed and no feed condition gives a lower bound on how much on-site sharing is due to interpersonal inuence along any communication medium. The mass adoption of online social networking systems has the potential to dramatically alter an individuals exposure to new information. By applying an experimental approach to measuring diusion outcomes within one of the largest human communication networks, we are able to rigorously quantify the eect of social networks on information spread. The present work sheds light on aggregate trends over a large population; future studies may investigate how properties of the individual, such as age, gender, and nationality, or features of content, such as popularity and breadth of appeal, relate to the inuence and its confounds.

comments messages photos threads

7.

ACKNOWLEDGMENTS

We would like to thank Michael D. Cohen, Dean Eckles, Emily Falk, James Fowler, and Brian Karrer for their discussions and feedback on this work. This work was supported in part by NSF IIS-0746646.

8.

REFERENCES

[1] L. A. Adamic and E. Adar. Friends and neighbors on the web. Social Networks, 25:211230, 2001. [2] E. Adar and A. Adamic, Lada. Tracking information epidemics in blogspace. In 2005 IEEE/WIC/ACM International Conference on Web Intelligence, Compiegne University of Technology, France, 2005. [3] E. Adar, J. Teevan, and S. T. Dumais. Resonance on the web: web dynamics and revisitation patterns. In Proceedings of the 27th International Conference on Human factors in Computing Systems, CHI 09, pages 13811390, New York, NY, USA, 2009. ACM Press. [4] A. Anagnostopoulos, R. Kumar, and M. Mahdian. Inuence and correlation in social networks. In Proceedings of the 14th Internal Conference on

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17] [18]

[19]

[20]

Knowledge Discover & Data Mining, pages 715, New York, NY, USA, 2008. ACM Press. S. Aral, L. Muchnik, and A. Sundararajan. Distinguishing inuence-based contagion from homophily-driven diusion in dynamic networks. Proc. Natl. Acad. Sci., 106(51):2154421549, December 2009. S. Aral and D. Walker. Creating social contagion through viral product design: A randomized trial of peer inuence in networks. Management Science, 57(9):16231639, Aug. 2011. L. Backstrom, D. Huttenlocher, J. Kleinberg, and X. Lan. Group formation in large social networks: membership, growth, and evolution. In KDD 06: Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 4454, New York, NY, USA, 2006. ACM. E. Bakshy, J. M. Hofman, W. A. Mason, and D. J. Watts. Everyones an inuencer: Quantifying inuence on twitter. In 3rd ACM Conference on Web Search and Data Mining, Hong Kong, 2011. ACM Press. E. Bakshy, B. Karrer, and L. Adamic. Social inuence and the diusion of user-created content. In Proceedings of the tenth ACM conference on Electronic commerce, pages 325334. ACM, 2009. H. R. Bernard, P. Killworth, D. Kronenfeld, and L. Sailer. The problem of informant accuracy: The validity of retrospective data. Annu. Rev. Anthropol., 13:495517, 1984. J. J. Brown and P. H. Reingen. Social ties and word-of-mouth referral behavior. J. Consumer Research, 14(3):pp. 350362, 1987. R. S. Burt. Structural holes: The social structure of competition. Harvard University Press, Cambridge, MA, 1992. D. Centola. The Spread of Behavior in an Online Social Network Experiment. Science, 329(5996):11941197, September 2010. D. Centola and M. Macy. Complex contagions and the weakness of long ties. Am. J. Sociol., 113(3):702734, Nov. 2007. M. Cha, A. Mislove, and K. P. Gummadi. A measurement-driven analysis of information propagation in the ickr social network. In Proceedings of the 18th international conference on World wide web, WWW 09, pages 721730, New York, NY, USA, 2009. ACM. N. A. A. Christakis and J. H. H. Fowler. The spread of obesity in a large social network over 32 years. N. Engl. J. Med., 357(4):370379, July 2007. S. Fox. The social life of health information. Technical report, Pew Internet & American Life Project, 2011. E. Gilbert and K. Karahalios. Predicting tie strength with social media. In Proceedings of the 27th International Conference on Human Factors in Computing Systems, CHI 09, pages 211220, New York, NY, USA, 2009. ACM. M. Gladwell. The Tipping Point: How Little Things Can Make a Big Dierence. Little Brown, New York, 2000. M. Gomez Rodriguez, J. Leskovec, and A. Krause.

[21]

[22] [23] [24]

[25]

[26]

[27]

[28]

[29]

[30]

[31]

[32]

[33]

[34]

[35] [36]

[37]

Inferring networks of diusion and inuence. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, KDD 10, pages 10191028, New York, NY, USA, 2010. ACM. A. Goyal, F. Bonchi, and L. V. Lakshmanan. Learning inuence probabilities in social networks. In Proceedings of the third ACM international conference on Web search and data mining, WSDM 10, pages 241250, New York, NY, USA, 2010. ACM. M. S. Granovetter. The strength of weak ties. Am. J. Sociol., 78(6):13601380, May 1973. M. S. Granovetter. Threshold models of collective behavior. Am. J. Sociol., 83(6):14201443, 1978. B. S. Greenberg. Person to person communication in the diusion of news events. Journalism Quarterly, 41:489494, 1964. D. Gruhl, R. Guha, D. Liben-Nowell, and A. Tomkins. Information diusion through blogspace. In Proceedings of the 13th international conference on World Wide Web, pages 491501. ACM, 2004. K. Hampton, L. S. Goulet, L. Rainie, and K. Purcell. Social networking sites and our lives. Technical report, Pew Internet & American Life Project, 2011. S. Hill, F. Provost, and C. Volinsky. Network-Based marketing: Identifying likely adopters via consumer networks. Stat. Sci., 21(2):256276, May 2006. G. Kossinets and D. J. Watts. Origins of homophily in an evolving social network. Am. J. Sociol., 115(2):405450, September 2009. K. Lerman and R. Ghosh. Information contagion: An empirical study of the spread of news on digg and twitter social networks. In Proceedings of 4th International Conference on Weblogs and Social Media (ICWSM), 2010. J. Leskovec, L. A. Adamic, and B. A. Huberman. The dynamics of viral marketing. In EC 06: Proceedings of the 7th ACM conference on Electronic commerce, pages 228237, New York, NY, USA, 2006. ACM. D. Liben-Nowell and J. Kleinberg. Tracing information ow on a global scale using internet chain-letter data. Proceedings of the National Academy of Sciences, 105(12):4633, 2008. C. F. Manski. Identication of endogenous social eects: The reection problem. Rev. Econ. Stud., 60(3):53142, July 1993. A. Marin. Are respondents more likely to list alters with certain characteristics? Implications for name generator data. Social Networks, 26(4):289307, Oct. 2004. M. McPherson, L. S. Lovin, and J. M. Cook. Birds of a Feather: Homophily in Social Networks. Annu. Rev. Sociol., 27(1):415444, 2001. M. E. J. Newman. Spread of epidemic disease on networks. Phys. Rev. E, 66(1):016128, Jul 2002. J.-P. Onnela and F. Reed-Tsochas. Spontaneous emergence of social inuence in online systems. Proceedings of the National Academy of Sciences, 107(43):1837518380, 2010. J. P. Onnela, J. Saramki, J. Hyvnen, G. Szab, a o o D. Lazer, K. Kaski, J. Kertsz, and A. L. Barabsi. e a

[38]

[39]

[40] [41]

[42]

[43]

Structure and tie strengths in mobile communication networks. Proceedings of the National Academy of Sciences, 104(18):73327336, May 2007. K. Purcell, L. Rainie, A. Mitchell, T. Rosenstiel, and K. Olmstead. Understanding the participatory news consumer. Technical report, Pew Internet & American Life Project, 2010. C. R. Shalizi and A. C. Thomas. Homophily and Contagion Are Generically Confounded in Observational Social Network Studies. Sociological Methods and Research, 27:211239, 2011. T. Stein, E. Chen, and K. Mangla. Facebook Immune System. In EuroSys Social Network Systems, 2011. E. S. Sun, I. Rosenn, C. A. Marlow, and T. M. Lento. Gesundheit! modeling contagion through facebook news feed. In Proceedings of the 3rd Intl AAAI Conference on Weblogs and Social Media, San Jose, CA, 2009. AAAI. D. J. Watts and S. H. Strogatz. Collective dynamics of small-world networks. Nature, 393(6684):440442, June 1998. S. Wu, J. M. Hofman, W. A. Mason, and D. J. Watts. Who says what to whom on twitter. In ACM Conference on the World Wide Web, Hyderbad, India, 2011. ACM Press.