Predicting online group opinions: A hypergraph-enhanced structure deep clustering with LSTM

Abstract

Understanding how group opinions form, shift, and polarize in online networks is critical for maintaining healthy public discourse and addressing the psychological drivers of digital behavior. While recent advances in computational modeling have improved prediction, most methods rely on pairwise graph structures that fail to capture higher-order dynamics and lack integration with behavioral theory.

To bridge this gap, we propose a psychologically grounded deep learning framework that combines hypergraph-enhanced structural clustering (HG-SDCN) with long short-term memory (LSTM) networks. Guided by Bandura's triadic reciprocal determinism, we construct a cognitive feature set encompassing environmental context, individual cognition, and behavioral expression—framing social behavior as an emergent property of cognitive–environmental interaction.

The HG-SDCN module models complex group relations through hypergraph convolution and dual self-supervision, yielding improved group detection. Subsequently, LSTM is used to capture temporal sentiment trajectories, outperforming traditional ARIMA in predictive accuracy.

Beyond prediction, our model offers conceptual insights into the formation and evolution of digital group cognition. By fusing psychological theory with deep learning, this interdisciplinary framework informs the design of socially aware AI systems, platform governance strategies, and interventions to counter online polarization.