Do Cascades Recur?

Cascades of information-sharing are a primary mechanism by which content reaches its audience on social media, and an active line of research has studied how such cascades, which form as content is reshared from person to person, develop and subside. In this paper, we perform a large-scale analysis of cascades on Facebook over significantly longer time scales, and find that a more complex picture emerges, in which many large cascades recur, exhibiting multiple bursts of popularity with periods of quiescence in between. We characterize recurrence by measuring the time elapsed between bursts, their overlap and proximity in the social network, and the diversity in the demographics of individuals participating in each peak. We discover that content virality, as revealed by its initial popularity, is a main driver of recurrence, with the availability of multiple copies of that content helping to spark new bursts. Still, beyond a certain popularity of content, the rate of recurrence drops as cascades start exhausting the population of interested individuals. We reproduce these observed patterns in a simple model of content recurrence simulated on a real social network. Using only characteristics of a cascade's initial burst, we demonstrate strong performance in predicting whether it will recur in the future.

Presented at WWW 2016.

Transcript

1. Do Cascades Recur? Justin Cheng, Lada Adamic, Jon Kleinberg, Jure Leskovec
3. Several weeks later…
12. Time Popularity Ahmed et al. (2013), Bauckhage et al. (2013), Matsubara et al. (2012), Yang & Counts (2010) Prior work: cascades rise, then fall
13. Mar Jun Sept Dec Cascades rise and fall many times Popularity
14. Mar Jun Sept Dec Cascades rise and fall many times Popularity
15. Mar Jun Sept Dec Cascades consist of multiple copies Popularity
16. Mar Jun Sept Dec Cascades consist of multiple copies Popularity
17. Mar Jun Sept Dec Cascades consist of multiple copies Popularity
18. Cascades are complex Mar Jun Sept Dec Popularity
19. Cascades are complex Time
20. Individual copies recur Time
21. Studying recurrence on Facebook 5 billion reshares 100 million photos 76 thousand clusters (Sampled over all of 2014 and de-identified)
22. 2 out of 5 image memes on Facebook recur.
23. 1 out of 3 videos on Facebook recur.
24. Cascades recur.
25. Cascades recur. Why do cascades recur? How do cascades recur? What is recurrence?
26. What is recurrence?
27. Defining recurrence Popularity r Time t
28. Defining recurrence p1 p4 p2 Popularity r Time t p3 (?)
29. Defining recurrence Popularity r Peaks have a minimum absolute/relative height rp ≥ h0, rp ≥ m·r *Not a peak Time t r p4 p1 p4 p2 pi pi rp p1 rp p2 rp p4
30. Defining recurrence Popularity r Peaks are local maxima rp ≥ max { rj | pi -w ≤ j ≤ pi +w } Time t p1 p4 p2 pi rp p1
31. Defining recurrence Popularity r Valleys separate peaks rp , rp ≥ v · max { rj | pi < j < pi+1 } Time t p1 p4 p2 pi pi+1 rp p1
32. Defining recurrence Popularity r Time t p1 p4 p2 Recurrences
33. Defining recurrence Popularity r Time t b1 b2 b4 Recurrences
34. Recurrence is common Number of Bursts Empirical CCDF 0.00 0.25 0.50 0.75 1.00 0 5 10 15
35. 0.00 0.25 0.50 0.75 1.00 0 5 10 15 Recurrence is common Number of Bursts Empirical CCDF 40% of cascades recur 0.40 2
36. 0.00 0.25 0.50 0.75 1.00 0 5 10 15 Recurrence is common Number of Bursts Empirical CCDF Cascades have 2.3 bursts on average 2.3
37. Recurrence takes time Days Between 1st and 2nd Burst Empirical CCDF 0.00 0.25 0.50 0.75 1.00 0 50 100 150 200
38. 0.00 0.25 0.50 0.75 1.00 0 50 100 150 200 Recurrence takes time Days Between 1st and 2nd Burst Empirical CCDF A meme takes an average of 32 days to recur 32
39. What is recurrence? A cascade recurs when it peaks in popularity more than once. Recurrence is common. Recurrence takes time.
40. How do cascades recur?
41. Do bursts occur in different parts of the network?
42. Bursts are (somewhat) separated Time Popularity
43. Bursts are (somewhat) separated Time Popularity
44. Bursts are (somewhat) separated Time Popularity
45. Bursts are (somewhat) separated Time Popularity 3.2 connections within bursts
46. Bursts are (somewhat) separated Time Popularity 1.4 connections across bursts
47. Are more popular cascades more likely to recur?
48. Moderate popularity increases recurrence Size of Initial Burst Mean Bursts 1.5 2.0 2.5 3.0 3.5 4.0 103 104 105 106
49. Large initial bursts exhaust susceptibles Size of Initial Burst Proportion Exposed in Second Burst 0.2 0.4 0.6 0.8 103 104 105 106
50. Are cascades with more diverse populations more likely to recur?
51. Moderate diversity increases recurrence Country Entropy in Initial Burst Mean Bursts Gender Entropy in Initial Burst 2.0 2.5 3.0 3.5 0 1 2 3 4 5 1.5 2.0 2.5 3.0 0.2 0.4 0.6 0.8 1.0
52. Do new copies of the same meme spark recurrence?
53. New copies can spark recurrence… Recurring cascades are spread out over more copies Recurring Non-recurring New copies correlate with recurrence r = 0.66 Introduction of new copies & number of reshares 4x
54. …but copies aren’t the only cause! Individual copies also recur 18% of individual copies recur Copies can be traced back to other copies 75% of copies can be attributed
 via the friendship graph
55. How do cascades recur? Bursts happen in separate parts of the network. Moderately viral/diverse content tends to recur. New copies can spark recurrence.
56. Why do cascades recur?
57. A model of recurrence
58. A model of recurrence 1
59. A model of recurrence 1 2
60. Low virality 1 2
61. Low virality 1 2
62. High virality 1 2
63. High virality 1 2
64. Moderate virality 1 2
65. Moderate virality 1 2
66. Moderate virality 1 2
67. Popularity Low virality Moderate virality High virality Overall Copy 1 Copy 2 Time
68. Popularity Low virality Moderate virality High virality Overall Copy 1 Copy 2 Time
69. Popularity Low virality Moderate virality High virality Overall Copy 1 Copy 2 Time
70. Popularity Low virality Moderate virality High virality Overall Copy 1 Copy 2 Time
71. Low virality Moderate virality High virality Popularity Recurrence No Recurrence No Recurrence Time Overall Copy 1 Copy 2
72. Simulations of this model replicate previous key findings.
73. Can we predict recurrence?
74. Features (e.g., burst length) (e.g., gender) (e.g., # edges) (e.g., # copies) Temporal Sharer Network Copy
75. Predicting recurrence Will it recur? Existence Will it be larger? Size When will it recur? Time
76. Predicting recurrence Will it recur? Will it be larger? When will it recur? Existence Size Time .89 .78 .58 AUC AUC AUC
77. Future / Related Work • Effect of Multiple Networks, External Stimuli
 Gruhl et al. (2004), Kumar et al. (2005), Myers & Leskovec (2012) • Improved Models of Recurrence
 Barabasi (2006), Cha et al. (2012), Matsubara et al. (2012) • Other Factors Influencing Recurrence (Seasonality, Sentimentality)
 Altizer et al. (2006), Verdasca et al. (2005)
78. Cascades Do Recur. Justin Cheng, Lada Adamic, Jon Kleinberg, Jure Leskovec http://bit.ly/cascades-paper