Source-linked AI summary
Predicting Successful Memes using Network and Community Structure
Lilian Weng, Filippo Menczer, Yong-Yeol Ahn
TL;DR
The paper asks whether successful memes can be predicted from their early spreading patterns. It develops and compares network, community, and temporal features in a machine-learning framework. Community-based features are the strongest predictors, while early popularity is not a reliable predictor of future popularity.
Problem
The paper investigates whether a meme's future popularity can be predicted from its early spreading patterns.
Method
The study compares features describing early-adopter influence, community structure, and adoption-time-series characteristics using machine-learning prediction models.
Results
Community-based features perform best, and the network-based approach outperforms baseline models, especially when predicting memes with few adopters.
Takeaways & Limitations
Early popularity alone is not a good predictor of future popularity, whereas community structure provides the strongest supported signal of meme success.
Takeaways & Limitations
Different approaches require different information: network-based features need network and adopter-position data, while timing models need early-message timestamps.
Abstract
from arXiv · showhide
We investigate the predictability of successful memes using their early spreading patterns in the underlying social networks. We propose and analyze a comprehensive set of features and develop an accurate model to predict future popularity of a meme given its early spreading patterns. Our paper provides the first comprehensive comparison of existing predictive frameworks. We categorize our features into three groups: influence of early adopters, community concentration, and characteristics of adoption time series. We find that features based on community structure are the most powerful predictors of future success. We also find that early popularity of a meme is not a good predictor of its future popularity, contrary to common belief. Our methods outperform other approaches, particularly in the task of detecting very popular or unpopular memes.
Introduction
The paper asks whether successful memes can be predicted from early spreading patterns and addresses the limited attention previous work gave to network structure. It compares audience, community, and temporal features, finding accurate early prediction and particularly important community-based signals.
- Only a few of the many memes created each day become viral, motivating early prediction of meme success.
- Existing prediction approaches analyze either early popularity time series or distinguishing features using supervised classification.
- The study compares features capturing early-adopter audience size, community structure, and early adoption growth.
- The model predicts meme popularity to an order of magnitude two months ahead using only a small number of early tweets, outperforming several baselines and regression models.
Background
Prior work links meme virality to content appeal, user behavior, social influence, network structure, and temporal adoption patterns. However, early popularity has been widely treated as a strong predictor, while the paper reports that initial breadth is a stronger indicator of larger cascades.
- Meme virality has been studied through innate attractiveness, influential users, and adoption patterns.
- Content-based studies associate virality with emotional, textual, and hashtag features, while randomized experiments suggest innate quality may play only a minor role under strong social influence.
- User interests, limited attention, and heterogeneous influence contribute to differences in meme popularity and longevity.
- Network structure affects spreading processes, including infection persistence, cascade-size distributions, and system vulnerability.
- Social contagion differs from disease transmission because multiple exposures can substantially increase adoption chances.
- Time-series analyses use growth, bursts, and decay to identify trendy or bursty memes, while classification studies often assume early popularity correlates strongly with future popularity.
- The results are consistent with the finding that initial breadth strongly indicates larger cascades.
Dataset
The study uses Twitter hashtags as identifiable memes and reconstructs their spreading environment from tweets and a reciprocal follower network. Community detection uses two contrasting algorithms to assess robustness.
- Twitter supplies meme content, network structure, spreading events, and user behavior data for studying meme diffusion.
- Each hashtag is treated as a meme because hashtags spread by imitation and can mutate, compete, and survive.
- Tweets were collected during March and April 2012, and the follower network was reconstructed using reciprocal follow links.
- Communities were identified with InfoMap and LinkClustering to test robustness across disjoint and overlapping community structures.
- Communities containing fewer than three nodes were excluded from the analysis.
Definitions
The paper formalizes hashtags as memes, defines popularity through tweets or adopters, and represents their early spread with adopter and timestamp sequences. It also illustrates how adopter positions determine the network surface and potential audience.
- A meme h is a hashtag whose tweets form T(h), while T_n(h) contains its earliest n tweets.
- A(h) is the set of all adopters of h, and A_n(h) contains users who posted at least one of its first n tweets.
- Meme popularity is measured by either the number of tweets |T(h)| or the number of adopters |A(h)|.
- The adopter sequence records the creator of each meme-containing tweet, allowing users to appear repeatedly when they tweet more than once.
- Tweet time series record each meme-containing tweet's timestamp, and T^τ(h) denotes tweets within duration τ from the first tweet.
- Network surfaces count directly exposed neighbors, so adopter degrees and positions determine potential audience size and its growth.
Characterizing Viral Memes
The paper characterizes viral memes through early network position, community diversity, and growth rate. Viral memes are associated with broader community spread and faster early diffusion, while slow early spreading corresponds to lower popularity.
- Prediction signals: The model uses network topology, community diversity, and growth rate to identify viral memes from early-stage information.These signatures are computed from early adopters and their spreading patterns.
- Network topology: Early adopters’ network positions reveal the size and growth of a meme’s potential audience.Network surface measures directly exposed neighbors, while distances between consecutive adopters estimate potential-audience growth over time.
- Community diversity: Community structure captures social reinforcement and homophily that can facilitate internal meme circulation.Dense within-community connectivity increases repeated exposure, while shared interests make community members more likely to adopt similar memes.
- Growth rate: The early spreading time is the interval between a meme’s first and n-th tweets and gauges its initial growth rate.Viral memes are expected to spread more quickly than others.
- Growth rate: Meme popularity significantly decreases when early spreading is slow, despite fluctuations when the early spreading interval is small.The relationship is shown for the interval between the first and 50th tweets.
Prediction Features
The prediction model combines 13 early-spreading features covering network topology, community concentration, and adoption timing. These features describe audience reach, adopter dispersion, community diversity, and the pace and variability of early adoption.
- Feature design: The model contains 13 features computed from the first n tweets of each hashtag, where n is small relative to final viral-hashtag activity.The three feature groups describe potential audience size, community diversity, and initial momentum.
- Visualization: Figure 2 visualizes viral and non-viral diffusion through early adopters and their same-community neighbors in a reciprocal-follow network.Users are nodes, reciprocal follows are links, and adopters are colored blue.
- Network features: Network features use early-adopter connectivity and position to estimate potential audience reach.They include early-adopter count, first- and second-surface sizes, average step distance, step-distance variability, and adopter diameter.
- Network features: The diameter feature measures the maximum network distance between any two early adopters and represents audience coverage.It is defined over adopters appearing within the first n tweets.
- Community features: Community features measure how widely early tweets and adopters are distributed across predefined communities.Entropy increases with diversity and decreases with concentration; intra-community interaction indicates the strength of community trapping.
- Growth-rate features: Time-series features use consecutive-tweet durations to measure average early spreading time and its relative variability.The mean and fluctuations of step durations form two prediction features.
Experiments
The experiments predict future meme popularity from early-observation features and compare the model with five baselines. Evaluation excludes content-based predictors because the study focuses on network spreading patterns.
- Experimental setup: The experiments predict the magnitude of future meme popularity from features calculated during early observation.Popularity is measured either by total tweets or by distinct adopters.
- Baselines: The study uses five baselines: random guessing, majority guessing, a social-influence model, and two regression models based on early popularity or growth sequences.The baselines cover trivial, influence-based, and time-series prediction strategies.
- Scope: Content-based prediction models are not considered because the evaluation uses only network spreading patterns.The comparison therefore focuses on information available from meme diffusion rather than meme content.
Network-based Prediction Model (Pn)
The model uses features computed from each meme’s first n tweets, while restricting the dataset to newly appearing hashtags and evaluating multiple observation thresholds.
- Model construction: Pn computes prediction features from each meme’s first n tweets, excluding hashtags with fewer than n tweets.The model uses 300 decision trees, each built from 5 randomly selected features.
- Evaluation setup: The analysis includes multiple values of n, with class membership counts reported separately for each observation threshold.Table 2 reports the number of new hashtags in each class for different n values.
- Data selection: The dataset is restricted to hashtags used during the first two weeks of March 2012 and appearing in fewer than 20 tweets during February.This filtering is intended to retain new memes; the same threshold is used for the baselines.
- Observation windows: Figure 4 examines how average tweet counts evolve until memes reach minimum thresholds of 25, 50, or 100 tweets.Dashed lines indicate when each threshold is reached.
- Class distribution: Only 48 memes reach the order of 10^4 tweets, while 33 reach the order of 10^4 adopters.These counts describe the sparsity of the highest-popularity classes.
Evaluation with F1 Score
The evaluation uses class-wise F1 scores because imbalanced classes make accuracy misleading. Network-based models perform best overall, especially for extreme popularity classes, while early popularity alone is weak for the most viral memes.
- Metric: F1 combines class-wise precision and recall, making it suitable for evaluating viral and non-viral classes separately under class imbalance.F1 ranges from 0 to 1 and is the harmonic mean of precision and recall.
- Early popularity: Early popularity does not guarantee future popularity for the most viral hashtags, and its correlation with final popularity is weak.Initially unpopular hashtags can later become popular, limiting the LN baseline’s performance.
- Class imbalance: Dominant classes receive good results from all models because class sizes are imbalanced, while the majority baseline has non-zero F1 only for the dominant class.This motivates reporting class-specific precision and recall rather than accuracy alone.
- Overall comparison: The network-based approach outperforms five baselines in most cases, especially for the most viral hashtags and hashtags with few adopters.In these extreme classes, all other baseline models can fail to classify any instances correctly.
- Feature groups: Community-based features generally produce the best results, particularly for detecting very popular memes.Combining all features gives Pn the best overall results.
- Evaluation design: Figure 5 reports F1 scores across usage and adopter-popularity classes for observation windows of 25, 50, and 100 tweets.The displayed results use Infomap communities; link clustering produces similar results.
- Rare-event detection: The network-based approach is strongest at detecting rare events, including extremely popular and extremely unpopular hashtags.The combined model outperforms all baselines for these rare-event cases.
Conclusion
The study finds that community-based features are the strongest predictors of meme virality, while early growth rate provides weaker predictive information. Its approach outperforms baselines and regression models, especially for minority popularity classes, but requires information whose availability determines the suitable prediction method.
- The prediction model outperforms baselines and regression models in most cases, especially for memes in crucial minority classes.The compared regression models cannot capture the most popular or most unpopular memes effectively.
- Community-based features perform best among the three feature classes for predicting meme virality.
- The network-based approach outperforms other baseline models when predicting adopter counts, particularly for memes with few adopters.Predicting the number of meme adopters is more difficult than predicting popularity more generally.
- Predictive performance improves with longer observation windows and remains robust across different community detection methods.
- Different prediction approaches require different information: network-based features need network positions, whereas timing features need timestamps of early messages.The available information therefore constrains which approach can be selected.
- Observing a few early messages to predict virality has potential applications in social media analytics, marketing, and advertising.