Social Reaction-Aware Heterogeneous Graph Modeling for Unseen Source-Group Fake News Detection

Abstract

: Existing fake news detection methods largely rely on single-source datasets, leading models to overfit platform-specific features and perform poorly on heterogeneous multi-source data. Even with the emergence of Large Language Models (LLMs), our benchmarks show that general-purpose LLMs still struggle to identify deceptive intent when source-specific context is unavailable. To address unseen-source-group generalization, we propose SHIELD (Social Heterogeneous Interaction Embedding for Latent Deception). SHIELD models interaction patterns shared across sources rather than relying only on isolated text features or semantic inference. Specifically, we construct a Social Reaction-Aware Heterogeneous Interaction Graph to capture consistencies and discrepancies between news claims and user reactions, supported by constrained semantic and stylistic anchors. We introduce a Hierarchical Attentive Aggregation mechanism to learn more transferable representations from structural patterns and sentiment feedback. Empirical results on the MCFEND benchmark, where G1 denotes diverse fact-checking sources, G2 denotes translated English fact-checking sources, and G3 denotes Weibo-source news, show that SHIELD remains competitive in mixed-source detection and achieves the highest Avg.F1 among the evaluated baselines on two controlled unseen-source-group subsets with less-skewed target-label distributions. Specifically, when trained on G3 and tested on G2-Controlled and G1-Controlled target subsets, SHIELD improves Avg.F1 by 2.19 and 3.26 percentage points over the strongest trained baseline, respectively. These results suggest that structural interaction indicators, supported by training-only semantic and stylistic anchors, can improve robustness under the evaluated source-group shifts.