Source-linked AI summary
Do Cascades Recur?
Justin Cheng, Lada A Adamic, Jon Kleinberg, Jure Leskovec
TL;DR
Long-term cascade behavior is less studied than the initial rise and fall, leaving open whether popular content can return after quiet periods. The paper analyzes Facebook cascades over longer timescales, characterizes recurrence, models it, and predicts it from initial-burst features. Recurrence is common, with virality and multiple copies associated with its likelihood, while very large initial bursts can inhibit later recurrence.
Problem
Prior temporal analyses focus on cascades’ initial rise and fall, while their behavior over longer timescales remains largely unexplored.
Source: §1 · p. 1Method
The authors analyze 2014 public Facebook photo and video reshares, characterize bursts and recurrence, model recurrence on a real network, and predict future recurrence from initial-burst features.
Source: §1 · pp. 2–3, §2.2 · p. 4Results
Recurrence is common, and moderately appealing cascades are most likely to recur; multiple copies help spark recurrence, while a large initial burst can inhibit it.
Source: §1 · p. 2, §7 · p. 10Takeaways & Limitations
The findings provide a large-scale account of content recurrence and suggest that virality and multiple copies are factors associated with recurring bursts.
Source: §1 · p. 2, §7 · p. 10The analysis covers a single social network, and understanding recurrence without multiple copies remains future work.
Source: §7 · p. 10
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
Long-term cascade behavior is largely unexplored: successful content may decline monotonically or return in renewed bursts after a lull. This study analyzes Facebook cascades over longer timescales, characterizes recurrence, and examines its drivers and predictability.
1. INTRODUCTION
- The study analyzes publicly viewable Facebook photo and video reshares from 2014 and defines recurrence by identifying multiple bursts in reshare activity.A burst is described as a spike in reshares over time; one burst indicates no recurrence.Source: §1 · p. 2
- 59% of image memes and 33% of videos recur; recurring bursts tend to be over a month apart, with little sustained resharing.Among large cascades, recurrence can follow a short or extended lull after activity falls to very low or zero levels.Source: §1 · pp. 1–2
- Moderately appealing cascades are most likely to recur: moderate initial reshare volume and moderate homophily or diversity correlate with higher recurrence.A large initial burst may inhibit later recurrence by effectively “immunizing” much of the susceptible population.Source: §1 · p. 2
- Multiple copies help content recur and spread to different network areas, while individual copies can also resurge; the model reproduces observed recurrence patterns.The paper reports individual-copy recurrence for 18% of image memes and 30% of videos.Source: §1 · p. 2
- Using features observed during the initial burst, recurrence prediction reaches ROC AUC 0.89 for image memes; recurrence timing is less predictable at 0.58.Content virality and multiple-copy features perform best; relative burst size is predicted at 0.78.Source: §1 · p. 3
2. TECHNICAL PRELIMINARIES
This section describes the Facebook reshare datasets and defines recurrence as multiple qualifying peaks in reshare activity. It also outlines the data scope and the operational rules used to identify bursts.
2.1 Dataset Description
- Researchers seeded image memes from March 2014's top 200,000 reshared images, clustered near-identical copies, and checked sampled clusters for clustering quality.They also filtered clusters with activity in October–December 2013 to identify content whose observed spread began in 2014, and studied top individual copies.Source: §2.1 · p. 3
- The 2014 samples include 105,198,380 images in 76,301 clusters and 6,748,622 videos in 156,145 clusters.The samples also cover hundreds of millions of users and pages and billions of reshares; deleted items were excluded.Source: §2.1 · p. 3
2.2 Defining Recurrence
- Table 1 reports recurrence in a large proportion of popular Facebook image memes and videos, distinguishing all cascades from those beginning in 2014.The caption does not provide the table's cell values.Source: Table 1 · p. 4
- Recurrence means multiple peaks in daily reshares, identified by peak-height and local-maximum rules plus a required dip between adjacent peaks.The operational settings were h0=10, m=2, w=7, and v=0.5; each peak corresponds to one burst.Source: §2.2 · p. 4
- The burst definition is rudimentary and does not require activity to reach zero between bursts; stricter definitions produced qualitatively similar findings.The authors note that future definitions could account more specifically for burst shape.Source: §2.2 · p. 4
3. CHARACTERIZING RECURRENCE
The section establishes that recurrence is common over long timescales and characterizes how its timing, participating users, network reach, and content copies relate to renewed bursts. Across these dimensions, moderate virality and diversity are associated with recurrence, while large initial bursts can limit later reach.
3.1 Recurrence is common
- 40% of image-meme cascades that began in 2014 recurred, and the mean interval between initial and later bursts was 32 days.More than 30% of recurring cascades resurfaced only after a month or more.Source: §3.1 · p. 4, Figure 5 · p. 5
3.2 Temporal patterns
- Cascades with larger, longer initial bursts are more likely to recur, but recurrence is greatest for moderately popular cascades, peaking at approximately 105 initial reshares for both image memes and videos.Recurring cascades start with 15,547 reshares and last 9.3 days, versus 6,128 reshares and 6.9 days for non-recurring cascades.Source: §3.2 · p. 4
- After an initial recurrence, the chance of another rises from 0.40 to 0.60, while later burst sizes shrink for less-recurring cascades and stay more stable for frequent ones.Subsequent bursts become briefer, and average lulls between bursts span 28–32 days for image memes and 30–44 days for videos.Source: §3.2 · p. 5, Figure 6 · p. 5
3.3 Sharer characteristics
- Moderate demographic diversity is associated with the most recurrence: country entropy peaks near 3.0, and higher gender entropy is also predictive, though more weakly.Country diversity predicts recurrence after controlling for initial-burst duration, peak height, and reshares; age effects vary between image memes and videos.Source: §3.3 · p. 5, Figure 7 · p. 6
3.4 Network structure
- The second burst can reach a different but partly overlapping population: recurring cascades have estimated initial reach of 8.2 million versus 4.0 million for non-recurring cascades.The second burst’s estimated potential reach averages 6.8 million; the initial and second exposed populations have Jaccard similarity 0.15.Source: §3.4 · p. 6
- Recurring bursts have little direct overlap, but remain connected across the network; 28% of people potentially exposed in the second peak were also exposed in the first.The Jaccard similarity of people across the first two bursts is 0.02, while their cross-burst network connections average 1.4 per person.Source: §3.4 · p. 6
3.5 Catalyzing recurrence
- Multiple copies are associated with recurrence and may catalyze new bursts, but they are neither necessary nor sufficient to explain it.Recurring cascades have more copies (2,277 versus 93), while only 6% of copies are reshared at least 10 times in a day and other cascade differences remain after controlling for copy counts.Source: §3.5 · p. 7
- Pages account for over 70% of reshares in the second burst of recurring cascades, while recurrence also occurs for individual copies, which recur 18% of the time.The most popular copy in a second burst is likely page-created; individual-copy recurrence averages 261 days with bursts 41 days apart.Source: §3.5 · p. 7
4. MODELING RECURRENCE
The authors explain recurrence through content virality and the availability of multiple copies, then test a simple recurrence model in simulations on a real social network. The simulations reproduce an interior maximum in recurrence and support susceptible-population depletion as a constraint at high virality.
4.1 Why do cascades recur?
- The proposed account makes virality primary: moderately popular, diverse cascades recur most, while large initial bursts can expose and inactivate people needed for later spread.Multiple copies can catalyze recurrence, but recurring and non-recurring cascades with similar copy counts still differ in virality.Source: §4.1 · p. 7
4.2 A simple model of recurrence
- The model predicts few successful spread attempts and a single small peak at low virality, observable recurrence at moderate virality, and inhibited later spread at high virality.At high virality, an early burst may reach a large portion of the susceptible population, making it harder for other copies to spread.Source: §4.2 · pp. 7–8, Figure 8 · p. 8
4.3 Simulating recurrence
- The simulation uses an SIR model with multiple outbreaks introduced at different times and lower-probability reinfection of resistant nodes.It assumes independent copies introduced over time, with target nodes sampled in proportion to degree; the authors note that single-copy recurrence and homophily also matter in reality.Source: §4.3 · p. 8
- For virality 6×10^-4≤pc_0<8×10^-4, simulations reproduced recurring cascades, with moderate virality yielding the most recurrence.At pc_0≥8×10^-4, large initial bursts suppressed subsequent bursts; the model also showed lower graph connectivity after those bursts (579 versus 1,065 for random node removal).Source: §4.3 · p. 8
5. PREDICTING RECURRENCE
The authors use only a cascade’s initial burst to predict whether it will recur, the size of its next burst, and its timing. Recurrence and relative burst size are more predictable than timing, and recurrence prediction remains strong for individual content copies.
5.1 Factors driving recurrence
- The feature set spans initial-burst timing and shape, demographics, network structure, and multiple-copy and page-related characteristics.The features include burst duration and gradients, demographic entropy, connections and potentially exposed population, and copy counts and reshares per copy.Source: §5.1 · p. 9
5.2 Does it recur?
- Recurrence prediction reached Accuracy=0.82, F1=0.81, and AUC=0.89; Table 2 summarizes strong recurrence and size prediction but weaker timing prediction.Multiple-copy features performed best overall, while temporal and network features also showed robust individual performance.Source: §5.2 · p. 9, Table 2 · p. 9
5.3 Will the recurrence be smaller/larger?
- Relative second-burst size was predicted with Acc=0.72, F1=0.69, AUC=0.78 for image memes, and AUC=0.85 for videos.Temporal features, especially a long initial burst, outperformed other feature sets.Source: §5.3 · p. 9
5.4 When does it recur?
- Recurrence timing was much less predictable: image memes had Acc=0.56, F1=0.51, AUC=0.58, and videos had AUC=0.60.Longer initial bursts were most indicative of earlier recurrence.Source: §5.4 · p. 9
5.5 Predicting recurrence for individual copies
- Individual-copy recurrence prediction also performed strongly, with Acc=0.80, F1=0.79, AUC=0.88 for image memes, and AUC=0.82 for videos.Network features were strongest, while longer initial bursts remained predictive; relative size and timing were also predictable for individual copies.Source: §5.5 · p. 10
6. RELATED WORK
Prior work studied diffusion, future popularity, burst detection, and epidemiological recurrence, while the paper examines longer-term dynamics of individual content on Facebook.
6. RELATED WORK
- Prior diffusion research examined how information spreads [7] [37], including meme similarity and competition for limited attention.Source: §6 · p. 10
- Popularity forecasting work modeled temporal diffusion or trends and often focused on an initial rise-and-fall pattern.This paper instead studies long-term content dynamics on Facebook.Source: §6 · p. 10
- Prior studies often observed multiple bursts for topics or hashtags, commonly attributing them to external stimuli [38], whereas this paper found recurrence even for an individual copy, with most recurrence aperiodic.The paper adapted an SIR model assuming multiple points of infection on a real social network, reproducing key observed recurrence characteristics.Source: §6 · p. 10
- The paper adopts a simple burst definition based on peaks relative to mean activity, among a broader literature on burst detection in streams and time series.Source: §6 · p. 10
7. DISCUSSION AND CONCLUSION
The analysis finds recurrence common and identifies initial-burst virality and multiple copies as contributors, while a simple model reproduces some observed patterns. Its scope is bounded by single-network analysis, and predicting when recurrence occurs remains challenging.
7. DISCUSSION AND CONCLUSION
- Long-lasting, moderately popular, moderately diverse initial bursts are most likely to recur; multiple copies spark recurrence, while homophily may also influence it.Source: §7 · p. 10
- The study analyzes one social network, and the association between multiple copies and recurrence does not hold for individual-copy recurrence.Understanding recurrence without multiple copies, including homophily's role, remains future work.Source: §7 · p. 10
- The simple model reproduces some observed features, including pronounced bursts separated by little activity and an internal maximum in burst count as reshares increase.Source: §7 · p. 10
- Predicting whether recurrence occurs is feasible, but predicting when remains challenging; sentimentality, humor, current events, and seasonality may add predictive factors.Source: §7 · p. 10