Predicting User Interaction Outcomes with Text and Interaction Graphs

Understanding user interaction outcomes in online communities is critical for maintaining healthy discussion environments. Existing approaches primarily rely on textual semantics while overlooking the relational structure of social interactions. In this paper, we propose a cross-modal framework that integrates signed interaction graphs with textual representations for predicting user interaction outcomes. Specifically, we construct positive (upvote) and negative (downvote) interaction graphs, and derive user reputation features based on graph-based propagation. These reputation features are then fused with contextualized textual representations derived from RoBERTa. Experimental results on real-world community data demonstrate that our approach significantly outperforms text-only baselines. The proposed method highlights the importance of modeling signed social structures and user representations in cross-modal social computing tasks.