Simple Adaptive Spectrum Graph Filters for Rumor Detection

Abstract

Rumor detection has become an important problem on social networks. Previous methods mostly employ deep learning approaches to extract rumor features and apply spatial graph convolutions to process rumor propagation trees. However, these methods face inherent issues in the spatial domain and struggle to handle global structures effectively. To address these problems, we propose a new detection model, Simple Adaptive Spectrum Graph Filters (SASGF). Specifically, we obtain the original parameterized graph filter bank through a truncated polynomial, which reduces the scale of model parameters. Next, we transform the scalar graph frequencies and filter sequences into rich vector representations via representation encoding. We then redefine the filters with an attention mechanism that leverages the positional information of the graph frequency sequence to obtain adaptive filter responses. Finally, we perform channel-oriented filtering on the graph signal spectrum. Experiments on two real-world datasets show that our method outperforms existing baselines.