A Multi-Graph Neural Network attention fusion framework for emotion-aware subgraph anomaly detection in social media fake news propagation

Fake news on social media threatens civic trust and public safety, with emotionally charged content accelerating its spread through retweets, replies, and shares. This study addresses the challenge of detecting anomalous propagation subgraphs that reflect coordinated misinformation campaigns. We propose the Multi-Graph Neural Network Attention-based Propagation Learning for Emotion-Aware Anomaly Detection framework (hereafter referred to as MAPLE). This framework integrates sentiment features from users’ historical posts with network structures, combining multiple Graph Neural Networks (Graph Convolutional Network, Graph Attention Network, and GraphSAGE) through an attention-based fusion mechanism. Subgraph embeddings are evaluated using a One-Class Support Vector Machine for anomaly detection. Theoretical analyses establish guarantees on mutual information preservation, variance reduction, anomaly margin amplification, and entropy maximization. Experiments on the PolitiFact and GossipCop datasets show that MAPLE consistently outperforms state-of-the-art baselines, improving F1-scores by 43% and 2.17% respectively, while maintaining robustness across datasets. Unlike prior works that treat structural or sentiment cues in isolation, MAPLE provides the first unified multi-Graph Neural Network fusion framework with emotional context and theoretical underpinnings for subgraph anomaly detection. • Introduces MAPLE: a novel multi-GNN fusion framework for fake news detection. • Fuses GCN, GAT, and GraphSAGE outputs via attention-based weighting. • Theoretical guarantees on mutual information, variance reduction, and generalization. • Outperforms state-of-the-art methods on PolitiFact and GossipCop datasets. • Validated through ablation studies and information-theoretic analysis.