Diseases, ideas, innovations, and behaviors spread through social networks^{1, 2, 3, 4, 5, 6, 7, 8, 9,10, 11, 12}. With the availability of large-scale, digitized data on social communication^13, 14, the study of diffusion of memes (units of transmissible information) has become feasible recently^{15, 16, 17, 18}. The questions of how memes spread and which will go viral have recently attracted much attention across disciplines, including marketing^6, 19, network science^20, 21, communication²², and social media analytics^23, 24, 25. Network structure can greatly affect the spreading process^15, 26, 27; for example, infections with small spreading rate persist in scale-free networks⁸. Existing research has attempted to characterize viral memes in terms of message content²², temporal variation^16, 24, influential users^19, 28, finite user attention^18, 21, and local neighborhood structure¹⁰. Yet, what determines the success of a meme and how a meme interacts with the underlying network structure is still elusive. A simple, popular approach in studying meme diffusion is to consider memes as diseases and apply epidemic models^3, 4. However, recent studies demonstrate that diseases and behaviors spread differently; they have therefore been referred to as simple versus complex contagions, respectively^9, 29.

Here we propose that network communities^30, 31, 32—strongly clustered groups of people—provide a unique vantage point to the challenge of predicting viral memes. We show that (i) communities allow us to estimate how much the spreading pattern of a meme deviates from that of infectious diseases; (ii) viral memes tend to spread like epidemics; and finally (iii) we can predict the virality of memes based on early spreading patterns in terms of community structure. We employ the popularity of a meme as an indicator of its virality; viral memes appear in a large number of messages and are adopted by many people.

Community structure has been shown to affect information diffusion, including global cascades^33, 34, the speed of propagation³⁵, and the activity of individuals^36, 37. One straight-forward effect is that communities are thought to be able to cripple the global spread because they act as traps for random flows^35, 36 (Fig. 1(A)). Yet, the causes and consequences of thetrapping effect have not been fully understood, particularly when structural trapping is combined with two important phenomena: social reinforcement and homophily. Complex contagions are sensitive to social reinforcement: each additional exposure significantly increases the chance of adoption. Although the notion is not new³⁸, it was only recently confirmed in a controlled experiment⁹. A few concentrated adoptions inside highly clustered communities can induce many multiple exposures (Fig. 1(B)). The adoption of memes within communities may also be affected by homophily, according to which social relationships are more likely to form between similar people^39, 40. Communities capture homophily as people sharing similar characteristics naturally establish more edges among them. Thus we expect similar tastes among community members, making people more susceptible to memes from peers in the same community (Fig. 1(C)). Straightforward examples of homophilous communities are those formed around language or culture (Fig. 1(D,E)); people are much more likely to propagate messages written in their mother tongue. Separating social contagion and homophily is difficult^41, 42, and we interpret complex contagion broadly to include homophily; we focus on how both social reinforcement and homophily effects collectively boost the trapping of memes within dense communities, not on the distinctions between them.

(A) Structural trapping: dense communities with few outgoing links naturally trap information flow. (B) Social reinforcement: people who have adopted a meme (black nodes) trigger multiple exposures to others (red nodes). In the presence of high clustering, any additional adoption is likely to produce more multiple exposures than in the case of low clustering, inducing cascades of additional adoptions. (C) Homophily: people in the same community (same color nodes) are more likely to be similar and to adopt the same ideas. (D) Diffusion structure based on retweets among Twitter users sharing the hashtag #USA. Blue nodes represent English users and red nodes are Arabic users. Node size and link weight are proportional to retweet activity. (E) Community structure among Twitter users sharing the hashtags #BBC and #FoxNews. Blue nodes represent #BBC users, red nodes are #FoxNews users, and users who have used both hashtags are green. Node size is proportional to usage (tweet) activity, links represent mutual following relations.

• Full size image (270 KB)

To examine and quantify the spreading patterns of memes, we analyze a dataset collected from Twitter, a micro-blogging platform that allows millions of people to broadcast short messages (‘tweets’). People can ‘follow’ others to receive their messages, forward (‘retweet’ or “RT” in short) tweets to their own followers, or mention (‘@’ in short) others in tweets. People often label tweets with topical keywords (‘hashtags’). We consider each hashtag as a meme.

Communities and communication volume

Do memes spread like complex contagions in general? If social reinforcement and homophily significantly influence the spread of memes, we expect more communication within than across communities. Let us define the weight w of an edge by the frequency of communication between the users connected by the edge. Nodes are partitioned into dense communities based on the structure of the network, but without knowledge of the weights (see Methods). For each community c, the average edge weights of intra- and inter-community links, w↻c and w↶c, quantify how much information flows within and across communities, respectively. We measure weights by aggregating all the meme spreading events in our data. If memes spread obliviously to community structure, like simple contagions, we would expect no difference between intra- and inter-community links. By contrast, we observe that the intra-community links carry more messages (Fig. 2(A)). Similar results have been reported from other datasets^35, 37. In addition, by defining the focus of an individual as the fraction of activity that is directed to each neighbor in the same community, f↻, or in different communities, f↶, we find that people interact more with members of the same community (Fig. 2(B)). All the results are statistically significant (  ) and robust across community detection methods (seeSupplementary Information for additional details).

We measure weights and focus in terms of retweets (RT) or mentions (@). We show (A)community edge weight and (B) user community focus using box plots. Boxes cover 50% of data and whisker cover 95%. The line and triangle in a box represent the median and mean, respectively.

• Full size image (78 KB)

Meme concentration in communities

These results suggest that communities strongly trap communication. To quantify this effect for individual memes, let us define the concentration of a meme in communities. We expect more concentrated communication and meme adoption within communities if the meme spreads like a complex contagion. To gauge this effect, we introduce four baseline models. The random sampling model (M1) assumes equal adoption probability for everyone, ignoring network topology and all activity. The simple cascade model (M2) simulates the spreading of simple contagions⁴³. The social reinforcement model (M3) employs a simple social reinforcement mechanism in addition to considering the network structure. In the homophily model (M4), users prefer to adopt the same ideas that are adopted by others in the same community. The simulation mechanisms of the four baseline models are summarized in Table 1.

We estimate the trapping effects on memes by comparing the empirical data with these models. Note that we only focus on new memes (see definition in Methods). Let us define the concentration of a meme h based on the proportions of tweets in each community. The usage-dominant community c^t(h) is the community generating most tweets with h. The usage dominance of h, r(h), is the proportion of tweets produced in the dominant community c^t(h) out of the total number of tweets T(h) containing the meme. We also compute the usage entropy H^t(h) based on how tweets containing h are distributed across different communities. The relative usage dominance  and entropy  are calculated using M1as baseline. Analogous concentration measures can be defined based on users. Let g(h) be the adoption dominance of h, i.e., the proportion of the U(h) adopters in the community with most adopters. The adoption entropy H^u(h) is computed based on how adopters of h are allocated across communities. The higher the dominance or the lower the entropy, the stronger the concentration of the meme. All measures are computed only based on tweets containing each meme in its early stage (first 50 tweets) to avoid any bias from the meme’s popularity.

Figures 3(A–D) demonstrate that non-viral memes exhibit concentration similar to (or stronger than) baselines M3 or M4, suggesting that these memes tend to spread like complex contagions. Note that models M2, M3, and M4 produce stronger concentration than random sampling (M1), because M2 incorporates the structural trapping effect in simple cascades, M3considers both structural trapping and social reinforcement, and M4 captures both structural trapping and homophily.

Changes in meme concentration as a function of meme popularity are illustrated by plotting relative (A) usage dominance, (B) adoption dominance, (C) usage entropy, and (D) adoption entropy. The relative dominance and entropy ratios are averaged across hashtags in each popularity bin, with popularity defined as number of tweets T or adopters U; error bars indicate standard errors. Gray areas represent the ranges of popularity in which actual data exhibit weaker concentration than both baseline models M3 and M4. The effect of multiple social reinforcement is estimated byaverage exposures for every meme. The exposures can be measured in terms of (E) tweets or (F) users. Similar results for different types of networks and community methods are described in SI.

• Full size image (296 KB)

Do all memes spread like complex contagions? While the majority of memes are not viral, viral memes are adopted differently. Their concentration in the empirical data is the same as that of the simple cascade model M2 (see the gray areas in Fig. 3(A–D)); community structure does not seem to trap successful memes as much as others. These memes spread like simple contagions, permeating through many communities.

Strength of social reinforcement

To further distinguish viral memes from others in terms of types of contagion, let us explicitly estimate the strength of social reinforcement. For a given meme h, we count the number of exposures that each adopter has experienced before the adoption and compute the average exposures across all adopters, representing the strength of social reinforcement on h, labelled as N(h). The exposures can be measured in terms of tweets N^t(h) or users N^u(h). We compute relative average exposures,  , using only tweets at the early stages (first 50 tweets). If this quantity is large, adoptions are more likely to happen with multiple social reinforcement and thus the meme spreads like a complex contagion. As shown in Fig. 3(E–F), viral memes require as little reinforcement as the simple cascade model M2, while non-viral memes need as many exposures as M3 or M4. We arrive at the same conclusion: viral memes spread like simple contagions rather than like complex ones.

Prediction

The above findings imply an intriguing possibility: high concentration of a meme would hint that the meme is only interesting to certain communities, while weak concentration would imply a universal appeal and therefore might be used to predict the virality of the meme. To illustrate this intuition about the predictive power of the community structure, we show in Fig. 4 how the diffusion pattern of a viral meme differs from that of a non-viral one, when analyzed through the lens of community concentration.

We represent each community as a node, whose size is proportional to the number of tweets produced by the community. The color of a community represents the time when the hashtag is first used in the community. (A) The evolution of a viral meme (#ThoughtsDuringSchool) from the early stage (30 tweets) to the late stage (200 tweets) of diffusion. (B) The evolution of a non-viral meme (#ProperBand) from the early stage to the final stage (65 tweets).

• Full size image (194 KB)

Let us therefore apply a machine learning technique, the random forests classification algorithm, to predict meme virality based on community concentration in the early diffusion stage. We employ two basic statistics based on early popularity and three types of community-based features in the prediction model, listed below.

1. Basic features based on early popularity. Two basic statistical features are included in the prediction model. The number of early adopters is the number of distinct users who generated the earliest tweets. The number of uninfected neighbors of early adopters characterizes the set of users who can adopt the meme during the next step.
2. Infected communities. The simplest feature related to communities is the number ofinfected communities, i.e., the number of communities containing early adopters.
3. Usage and adoption entropy. H^t(h) and H^u(h) are good indicators of the strength of meme concentration, as shown in Fig. 3.
4. Fraction of intra-community user interactions. We count pair-wise user interactions about any given meme, and calculate the proportion that occur between people in the same community.

Our method aims to discover viral memes. To label viral memes, we rank all memes in our dataset based on numbers of tweets or adopters, and define a percentile threshold. A threshold of θT or θU means that a meme is deemed viral if it is mentioned in more tweets than θT% of the memes, or adopted by more users than θU% of the memes, respectively. All the features are computed based on the first 50 tweets for each hashtag h. Two baselines are set up for comparison. Random guess selects nviral memes at random, where nviral is the number of viral memes in the actual data. Community-blind prediction employs the same learning algorithm as ours but without the community-based features. We compute both precision and recall for evaluation; the former measures the proportion of predicted viral memes that are actually viral in the real data, and the latter quantifies how many of the viral memes are correctly predicted. Our community-based prediction excels in both precision and recall, indicating that communities are helpful in capturing viral memes (Fig. 5). For example, when detecting the most viral memes by users (θU = 90), our method is about seven times as precise as random guess and over three times as precise as prediction without community features. We achieve a recall over 350% better than random guess and over 200% better than community-blind prediction. Similar results are obtained using different community detection methods or different types of social network links (see SI).

We predict whether a meme will go viral or not; a meme is labeled as viral if it produces more tweets (T) or is adopted by more users (U) than a certain percentile threshold (θ = 70, 80, 90) of memes. We use the random forests classifier trained on community concentration features, which are calculated based on the initial n = 50 tweets for each meme. Prediction results are robust across different networks and community detection methods (see SI). We compute precision and recall to compare our prediction results against two baselines.

• Full size image (132 KB)

Despite the vast and growing literature on network communities, the importance of community structure has not been fully explored and understood. Our findings expose an important role of community structure in the spreading of memes. While the role of weak ties between different communities in information diffusion has been recognized for decades^35, 36, we provide a direct approach for translating data about community structure into predictive knowledge about what information will spread virally. Our method does not exploit message content, and can be easily applied to any socio-technical network from a small sample of data. This result can be relevant for online marketing and other social media applications.

Further analyses of network community structure in relation to social processes hold potential for characterizing and forecasting social behavior. We believe that many other complex dynamics of human society, from ethnic tension to global conflicts, and from grassroots social movements to political campaigns^17, 44, 45, could be better understood by continued investigation of network structure.

We collected a 10% sample of all public tweets from Mar 24 to Apr 25, 2012 using the Twitter streaming API (dev.twitter.com/docs/streaming-apis). Only tweets written in English are extracted. The dataset comprises 121,807,378 tweets generated by 14,599,240 unique users, and containing at least one of 10,393,465 hashtags. We then constructed an undirected, unweighted network based on reciprocal following relationships between 595,460 randomly selected users, as bi-directional links reflect more stable and reliable social connections. Such a conservative choice to exclude information about direction and weights of links makes the approach more generally applicable to cases where static data about the social network is more readily available than dynamic data about information flow. Two other types of networks constructed on the basis of retweets and mentions were also tested for robustness (see extended analyses in SI).

We apply Infomap³¹, an established algorithm to identify the community structure. To ensure the robustness of our results, we perform the same analyses using another widely-used but very different community detection method, link clustering³². The results are similar (see details in SI). The network remains unweighted for community identification, to focus purely on the connection structure.

For quantifying meme concentration in communities and the strength of social reinforcement, we focus on new memes that emerged during our observation time window. New memes are defined as those with fewer than 20 tweets during the previous month (Feb 24 – Mar 23, 2012). A sensitivity test of our results with respect to hashtag filtering criteria is available in SI.

To replicate the Twitter API sampling effect in the baseline models, each simulation runs until 10 times more tweets are generated than the empirical numbers. Then, we select 10% of the tweets at random. Every simulation is repeated 100 times and the 10%-sampling is repeated 10 times on each simulation outcome. Thus, the average values of the measures from our toy models are computed across 100 × 10 samples.

In prediction, we use the random forest algorithm, an ensemble classifier that constructs 500 decision trees⁴⁶. Each decision tree is trained with 4 random features independently and the final prediction outcomes combine the outputs of all the trees. The good performance of the random forest model benefits from the assumption that an ensemble of “weak learners” can form a “strong learner.” For training and testing, we employ 10-fold cross validation.

1. Newman, M. E. J. Networks: An Introduction (Oxford University Press, 2010).
   • Show context
2. Easley, D. & Kleinberg, J. Networks, Crowds, and Markets: Reasoning About a Highly Connected World (Cambridge University Press, 2010).
   • Show context
3. Goffman, W. & Newill, V. A. Generalization of epidemic theory: An application to the transmission of ideas. Nature 204, 225–228 (1964).
   • CAS
   • ADS
   • ISI
   • PubMed
   • Article
   • Show context
4. Daley, D. J. & Kendall, D. G. Epidemics and rumours. Nature 204, 1118–1118 (1964).
   • CAS
   • ADS
   • PubMed
   • Article
   • Show context
5. Anderson, R. M., May, R. M. & Anderson, B. Infectious Diseases of Humans: Dynamics and Control (Oxford Science Publications, 1992).
   • Show context
6. Rogers, E. Diffusion of Innovations (Free Press, 2003), 5th edn.
   • Show context
7. Barrat, A., Barthélemy, M. & Vespignani, A. Dynamical Processes on Complex Networks(Cambridge University Press, 2008).
   • Show context
8. Pastor-Satorras, R. & Vespignani, A. Epidemic spreading in scale-free networks. Phys. Rev. Lett. 86, 3200–3203 (2001).
   • CAS
   • ADS
   • ISI
   • PubMed
   • Article
   • Show context
9. Centola, D. The spread of behavior in an online social network experiment. Science 329,1194–1197 (2010).
   • CAS
   • ADS
   • PubMed
   • Article
   • Show context
10. Ugander, J., Backstrom, L., Marlow, C. & Kleinberg, J. Structural diversity in social contagion. PNAS 109, 5962–5966 (2012).
    • ADS
    • PubMed
    • Article
    • Show context
11. Bond, R. M. et al. A 61-million-person experiment in social influence and political mobilization. Nature 489, 295–298 (2012).
    • CAS
    • ADS
    • PubMed
    • Article
    • Show context
12. Christakis, N. A. & Fowler, J. H. The spread of obesity in a large social network over 32 years. New England Journal of Medicine 357, 370–379 (2007).
    • CAS
    • PubMed
    • Article
    • Show context
13. Vespignani, A. Predicting the behavior of techno-social systems. Science 325, 425–428(2009).
    • CAS
    • ADS
    • PubMed
    • Article
    • Show context
14. Lazer, D. et al. Computational social science. Science 323, 721–723 (2009).
    • CAS
    • PubMed
    • Article
    • Show context
15. Moreno, Y., Nekovee, M. & Pacheco, A. F. Dynamics of rumor spreading in complex networks. Physical Review E 69, 066130 (2004).
    • CAS
    • ADS
    • Article
    • Show context
16. Leskovec, J., Backstrom, L. & Kleinberg, J. Meme-tracking and the dynamics of the news cycle. In: Proc. ACM SIGKDD Intl. Conf. on Knowledge discovery and data mining,497–506 (2009).
    • Show context
17. Borge-Holthoefer, J. et al. Structural and dynamical patterns on online social networks: the spanish may 15th movement as a case study. PLOS ONE 6, e23883 (2011).
    • CAS
    • PubMed
    • Article
    • Show context
18. Lehmann, J., Gonçalves, B., Ramasco, J. J. & Cattuto, C. Dynamical classes of collective attention in twitter. In: Proceedings of the 21st international conference on World Wide Web, 251–260 (ACM, 2012).
    • Show context
19. Aral, S. & Walker, D. Creating social contagion through viral product design: A randomized trial of peer influence in networks. Management Science 57, 1623–1639(2011).
    • Article
    • Show context
20. Leskovec, J., Adamic, L. & Huberman, B. The dynamics of viral marketing. ACM Trans. Web 1 (2007).
    • Show context
21. Weng, L., Flammini, A., Vespignani, A. & Menczer, F. Competition among memes in a world with limited attention. Sci. Rep. 2, 335 (2012).
    • CAS
    • ADS
    • PubMed
    • Article
    • Show context
22. Berger, J. & Milkman, K. L. What makes online content viral? Journal of Marketing Research 49, 192–205 (2009).
    • Article
    • Show context
23. Jamali, S. & Rangwala, H. Digging digg: Comment mining, popularity prediction, and social network analysis. In: Proc. Intl. Conf. on Web Information Systems and Mining (WISM), 32–38 (2009).
    • Show context
24. Szabo, G. & Huberman, B. A. Predicting the popularity of online content.Communications of the ACM 53, 80–88 (2010).
    • Article
    • Show context
25. Suh, B., Hong, L., Pirolli, P. & Chi, E. H. Want to be retweeted? large scale analytics on factors impacting retweet in twitter network. In Proc. IEEE Intl. Conf. on Social Computing, 177–184 (2010).
    • Show context
26. Watts, D. J. A simple model of global cascades on random networks. PNAS 99,5766–5771 (2002).
    • CAS
    • ADS
    • PubMed
    • Article
    • Show context
27. Dodds, P. S. & Watts, D. J. Universal behavior in a generalized model of contagion.Phys. Rev. Lett. 92, 218701 (2004).
    • CAS
    • PubMed
    • Article
    • Show context
28. Kitsak, M. et al. Identification of influential spreaders in complex networks. Nature Physics 6, 888–893 (2010).
    • CAS
    • ADS
    • Article
    • Show context
29. Backstrom, L., Huttenlocher, D., Kleinberg, J. & Lan, X. Group formation in large social networks: membership, growth, and evolution. In: Proc. ACM SIGKDD Intl. Conf. on Knowledge discovery and data mining, 44–54 (2006).
    • Show context
30. Fortunato, S. Community detection in graphs. Physics Reports 486, 75–174 (2010).
    • ADS
    • Article
    • Show context
31. Rosvall, M. & Bergstrom, C. T. Maps of random walks on complex networks reveal community structure. PNAS 105, 1118–1123 (2008).
    • ADS
    • PubMed
    • Article
    • Show context
32. Ahn, Y.-Y., Bagrow, J. & Lehmann, S. Link communities reveal multiscale complexity in networks. Nature 466, 761–764 (2010).
    • CAS
    • ADS
    • ISI
    • PubMed
    • Article
    • Show context
33. Galstyan, A. & Cohen, P. Cascading dynamics in modular networks. Physical Review E75, 036109 (2007).
    • CAS
    • ADS
    • Article
    • Show context
34. Gleeson, J. P. Cascades on correlated and modular random networks. Physical Review E 77, 046117 (2008).
    • CAS
    • ADS
    • Article
    • Show context
35. Onnela, J.-P. et al. Structure and tie strengths in mobile communication networks. PNAS104, 7332–7336 (2007).
    • CAS
    • ADS
    • PubMed
    • Article
    • Show context
36. Granovetter, M. S. The strength of weak ties. American Journal of Sociology 78,1360–1379 (1973).
    • ISI
    • Article
    • Show context
37. Grabowicz, P. A., Ramasco, J. J., Moro, E., Pujol, J. M. & Eguiluz, V. M. Social features of online networks: The strength of intermediary ties in online social media. PLOS ONE7, e29358 (2012).
    • CAS
    • PubMed
    • Article
    • Show context
38. Granovetter, M. S. Threshold models of collective behavior. American Journal of Sociology 83, 1420–1433 (1978).
    • Article
    • Show context
39. McPherson, M., Lovin, L. & Cook, J. Birds of a feather: Homophily in social networks.Annual Review of Sociology 27, 415–444 (2001).
    • ISI
    • Article
    • Show context
40. Centola, D. An experimental study of homophily in the adoption of health behavior.Science 334, 1269–1272 (2011).
    • CAS
    • ADS
    • PubMed
    • Article
    • Show context
41. Aral, S., Muchnik, L. & Sundararajan, A. Distinguishing influence-based contagion from homophily-driven diffusion in dynamic networks. PNAS 106, 21544–21549 (2009).
    • ADS
    • PubMed
    • Article
    • Show context
42. Shalizi, C. & Thomas, A. Homophily and contagion are generically confounded in observational social network studies. Sociological Methods & Research 40, 211–239(2011).
    • PubMed
    • Article
    • Show context
43. Karsai, M. et al. Small but slow world: How network topology and burstiness slow down spreading. Physical Review E 83, 025102 (2011).
    • CAS
    • ADS
    • Article
    • Show context
44. Conover, M. D., Ferrara, E., Menczer, F. & Flammini, A. The digital evolution of occupy wall street. PLOS ONE 8, e64679 (2013).
    • CAS
    • ADS
    • PubMed
    • Article
    • Show context
45. Conover, M. D. et al. The geospatial characteristics of a social movement communication network. PLOS ONE 8, e55957 (2013).
    • CAS
    • ADS
    • PubMed
    • Article
    • Show context
46. Breiman, L. Random forests. Machine Learning 45, 5–32 (2001).
    • ISI
    • Article
    • Show context
47. Romero, D. M., Tan, C. & Ugander, J. On the interplay between social and topical structure. In: Proc. AAAI Intl. Conf. on Weblogs and Social Media (2013).
    • Show context

Download references

We thank Albert-László Barabási, Haewook Kwak, James P. Bagrow, Cosma Rohilla Shalizi, Yang-Yu Liu, Sune Lehmann, and Sue Moon for helpful discussions and comments on earlier versions of this manuscript, and Yangyi Chen for help on using FutureGrid clusters for processing large-scale data. This manuscript is based upon work supported in part by the James S. McDonnell Foundation and the National Science Foundation under Grants No. 0910812 and 1101743. While finalizing this manuscript, we have been made aware of a related paper developed independently by D.R. Romero, C. Tan, and J. Ugander⁴⁷.