Graph-Based Methods¶
Graph-based methods treat social media users and their interactions as networks, where nodes represent users and edges represent relationships or communication patterns (follows, mentions, retweets). These approaches extract structural features—clustering, centrality, modularity, random walks—to infer properties like authenticity, influence, polarization, and controversy.
Key finding¶
Network topology often encodes phenomena that content analysis alone cannot capture: graph structure (clustering, boundary connectivity, random walk properties) reliably identifies controversy, reveals echo chambers, and distinguishes coordinated from organic behavior.
Key papers¶
- Shu et al. (2019) — Hierarchical Propagation Networks for Fake News Detection: Models news propagation as hierarchical directed graphs at two levels of granularity: macro-level retweet cascades (nodes=tweets, edges=retweets) and micro-level reply trees (nodes=replies, edges=reply-to relationships); extracts graph features including structural properties (depth, outdegree, connectivity patterns) and bot presence; achieves F1 > 0.80 on FakeNewsNet showing that hierarchical graph structure encodes veracity signals.
- Garimella et al. (2017) — Quantifying Controversy on Social Media — Proposes multiple graph-based measures (random walk, betweenness centrality, embedding, boundary connectivity, dipole moment) for controversy quantification; shows random-walk-based metric most reliably separates controversial from non-controversial topics.
- Garimella et al. (2016) — Reducing Controversy by Connecting Opposing Views — Applies random-walk-based controversy metric to network optimization; develops efficient graph algorithm for edge-recommendation problem that selectively adds connections between high-degree nodes to minimize controversy.
- Cinelli et al. (2021) — The echo chamber effect on social media — Uses network clustering and modularity analysis to characterize echo chambers across platforms.