Heterogeneous Graph Distillation for Stance Prediction

Stance prediction is a critical task in public opinion analysis, aiming to identify users' viewpoints on specific events. Existing research often relies on user interactions for stance inference but generally underutilizes multi-source heterogeneous information such as user entities, opinion text, issues and topics. To address this limitation, this study proposes a stance prediction approach based on heterogeneous entity modeling. By integrating four types of heterogeneous entities to capture similarity in users' participation in issues, the proposed method improves stance inference accuracy. Specifically, we design a heterogeneous graph knowledge extraction framework that fully incorporates both content features and structural semantic information of various entities. First, we construct a heterogeneous information network to capture different types of social media entities and their interactions, learning rich feature representations in the process. Next, we employ matrix factorization to assess users' preferences toward specific issues. Finally, by introducing a knowledge distillation mechanism, the approach significantly enhances prediction accuracy with only a modest increase in computational cost. Experimental results on public datasets demonstrate that our method outperforms existing baselines, verifying its effectiveness.