SocioHGN: A Heterogeneous Graph Networks Augmented by Social Mechanism

Analyzing online opinion dynamics during controversial global events presents significant challenges. While existing graph learning models identify structural patterns, they overlook key social mechanisms like collective meaning formation and emotional contagion that shape social relationships and community evolution. We propose SocioHGN, an innovative model integrating sociological theory to simulate interactions among users, topics, and geographic locations. Our approach employs frame-aware subgraph extraction combined with graph attention networks to capture complex social dynamics. We validate our model using Weibo data from the Russia-Ukraine conflict, comparing against established graph neural network baselines. Results show SocioHGN significantly outperforms existing methods in link prediction (AUC) and social consistency metrics (RAS, EDI). Moreover, SocioHGN provides interpretable insights, quantitatively revealing how frame evolution drives community polarization, demonstrating its value for both prediction and social analysis.