Spectral graph filter bank fusion for multimodal rumor cascade representation and detection

Rumor detection on social media is inherently a multi-source information fusion problem that involves heterogeneous signals from textual content, user interactions, and temporal dynamics. Existing approaches often underexploit the global structural patterns encoded in rumor propagation graphs. In this work, we propose a unified multimodal fusion framework, termed Rumor Graph Filter Net (RGFN), which integrates multimodal representations with learnable spectral graph filtering. Specifically, post content embeddings derived from a pre-trained language encoder, user social relation representations obtained via self-supervised contrastive learning, and propagation timestamps are jointly transformed into the graph spectral domain using the Graph Fourier Transform (GFT). To adaptively capture discriminative propagation characteristics, we design channel-wise graph filter banks based on Chebyshev polynomial approximations, enabling frequency-selective enhancement and noise suppression. The filtered spectral responses are projected back to the spatial domain through the inverse GFT, followed by a simplified hierarchical multi-view graph pooling mechanism to obtain robust graph-level representations across scales. The final classifier operates on fused multilevel features for rumor identification. Experiments on two real-world rumor datasets, Twitter and Weibo, show that RGFN achieves the highest F1-score among the baselines and performs competitively on other metrics. Specifically, RGFN outperforms the best baselines on Twitter and Weibo by 4.9% and 1.4% on F1-score, respectively. The proposed framework offers a principled integration of graph signal processing and multimodal information fusion for social media credibility assessment.