Cascade dynamics and prediction

Cascade dynamics examines the structural and temporal properties of information cascades—trees of shares, retweets, and reposts that arise from an initial post. Research addresses both understanding cascade properties (depth, breadth, growth patterns) and predicting cascade evolution from partial observations.

Core questions: How do cascades grow over time? What fraction of potential reachable users will share? How do structural properties of networks and early cascade characteristics predict final cascade size?

Temporal dynamics: Early observations are often predictive of final size, but this relationship varies by platform and content type. Some cascades explode quickly; others accumulate shares slowly over long periods. The temporal patterns of sharing (burstiness, interarrival times) carry information about cascade trajectory.

Structural properties: Cascade depth (longest path from origin to leaf) and breadth (number of reshares at each level) distinguish different propagation mechanisms. Rumor cascades, for example, tend to run deeper than typical content, indicating spread via weak ties rather than seed account influence.

Prediction challenges: Early cascade features (number of shares in first hour, diversity of initial sharers, temporal gaps) must predict final size without knowledge of intervening period. Deep learning approaches learn cascade embeddings; RNNs and graph neural networks capture both temporal and topological patterns.

Key papers

• The structure and function of complex networks — Foundational review of network science; provides mathematical framework for understanding spreading processes on networks
• A Survey of Information Cascade Analysis: Models, Predictions, and Recent Advances — Comprehensive survey of cascade prediction; covers 250+ papers; systematic taxonomy of features (temporal, structural, user, content); compares feature-based, generative, and deep learning approaches
• Rumor Cascades — Analyzes rumor cascade depth distribution; shows rumor cascades run deeper than reference photo cascades of similar size, indicating content-driven rather than seed-driven spread
• Can Cascades be Predicted? — Early prediction of cascade growth from temporal patterns; identifies temporal features as most predictive (~80% accuracy)
• DeepCas: an End-to-end Predictor of Information Cascades — Deep learning cascade prediction using graph embeddings and random walks
• Topological Recurrent Neural Network for Diffusion Prediction — Topological LSTM for cascade prediction accounting for network structure

Connections

• Information diffusion in social networks — broader diffusion processes
• Rumor propagation on social networks — cascade dynamics of rumors specifically
• Social networks and online communities — network topology enabling cascades
• Feature engineering for fake news detection — features extracted from partial cascades for prediction