StyleGraph: A Heterogeneous Graph Neural Framework for Stylistic and Semantic Rumor Detection on Social Media

Identifying and controlling misinformation circulated on social media is known as rumor detection. The growing prevalence of deceptive content presents a significant challenge that necessitates advanced detection techniques, beyond classical machine learning. This study introduced StyleGraph, a novel graph-based framework that enhances rumor detection by jointly leveraging semantic and stylistic information. Unlike previous methods that focused solely on textual content or propagation structures, Style-Graph constructs a heterogeneous graph that models interactions among authors, phrases, and comments. We integrate stylistic features derived from author-writing patterns with semantic representations to enrich node embeddings. StyleGraph was evaluated using Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), and Cluster GCNs on the PHEME dataset under both 5-event and 9-event settings. The Cluster GCN model achieved the highest accuracy of 62.7% and an F1-score of 55.5% on the 9-event dataset, while the GAT model achieved 77.7% accuracy on the 5-event dataset. Further improvements were observed with the use of data augmentation and sentence-BERT (SBERT) embeddings, with GAT achieving an F1-score of 75.3%. The proposed framework demonstrates scalability and effectiveness in real-time rumor detection scenarios.