Fuzzy relation-aware graph role clustering for sentiment-augmented fake news detection

Existing research on fake news detection shows that sentiment information in social media comments can provide important clues for judging the authenticity of news. However, existing sentiment enhancement methods usually treat comments as independent samples and mainly rely on the semantic representation and sentiment features of individual comments for role clustering, ignoring the potential relationship structure between comments. In fact, the user roles in the comment section are not entirely determined by the sentiment polarity of individual comments. Comments expressing the same negative sentiment may come from blind followers, skeptics, or debunkers. To address this issue, we propose a fuzzy relation-aware graph role clustering framework, FRGC. Specifically, FRGC models comments under the same news event as nodes and constructs an event-level comment graph based on semantic similarity to capture latent comment relations beyond individual representations. It further introduces sentiment probability distributions and their uncertainty to calibrate semantic edge weights, avoiding the direct use of unreliable sentiment polarity for relation modeling. Through this design, sentiment information is incorporated as an uncertain auxiliary signal while preserving the dominance of semantic relations. Furthermore, a graph encoder layer and a weak graph regularization term are introduced to obtain relation-enhanced comment representations and improve the stability of latent role clustering. Experimental results on multiple datasets validate the effectiveness of FRGC. Specifically, FRGC achieves an F1-score of 97.1 on Weibo-comp and obtains a Macro-F1 of 36.5 and RMSE of 74.3 on RumourEval-19, outperforming existing methods and demonstrating the effectiveness of fuzzy relation-aware graph modeling for sentiment-enhanced fake news detection.