Distilling the Noise: Value-Aware Contrastive Hypergraph Learning for Robust Multimodal Rumor Detection

Abstract

The proliferation of multimodal misinformation on social media poses significant challenges for rumor detection. Existing approaches typically treat user comments as homogeneous auxiliary features for aggregation. However, this assumption overlooks the significant heterogeneity in comment quality, where high-value comments containing logical reasoning are often submerged by varying degrees of emotional noise or coordinated manipulation, thereby degrading detection performance. To address the underutilization of comment signals and the poor signal-to-noise ratio, this paper proposes a Multiview Hypergraph-based Multimodal Rumor Detection (MHMRD) framework. Specifically, the framework operates in two key phases. First, we introduce a Value-Aware Comment Discrimination Module, which leverages Large Language Models (LLMs) to construct value-aware supervision, training a discriminator to suppress low-utility noise while amplifying credible evidence. Second, we construct a Multi-Scale Hypergraph Network encompassing micro-scale (intra-modal), meso-scale (comment interaction), and macro-scale (inter-post) views to capture high-order semantic correlations. Finally, these heterogeneous views are jointly optimized via a Dual Contrastive Learning strategy to achieve deep feature alignment. Extensive experiments on the Weibo and PHEME datasets demonstrate that MHMRD significantly outperforms state-of-the-art baselines, validating its effectiveness and robustness in complex social contexts.