Semantic-Aware Heterogeneous Graph Learning for Fake News Detection

While heterogeneous graph learning (HGL) is a promising approach for fake news detection, current methods struggle to capture complex relationships among neighborhoods due to their reliance on binary interactions. This limitation can lead to missed critical information and jeopardize performance. We address this issue by proposing a novel semantic-aware heterogeneous graph representation learning (SHEAR) method. SHEAR leverages network motifs, i.e., higher order substructures, to represent heterogeneous behavioral semantic patterns, which are then encoded into low-dimensional embeddings. The semantic information is incorporated into the news nodes by aggregating motif instance embeddings of the same type via an instance-level attention module. SHEAR assigns scores to each motif type using a semantic-level attention module, reflecting the relative importance of different semantic information. The final news node representations are obtained by fusing these embeddings. Experimental results on real-world datasets demonstrate that SHEAR significantly outperforms existing methods in fake news detection, showcasing the benefits of incorporating behavioral semantic patterns with network motifs for enhanced classification performance.