RHEA Enhances Multimodal Graph Clustering with Reliability-Awareness

A novel framework named RHEA (Reliability-Harmonized Reconstruction and Assignment) has been introduced to improve clustering in multimodal-attributed graphs (MAGs). MAGs are complex data structures where nodes possess diverse attributes, such as text and images, and are connected by relational links. Existing clustering methods for MAGs often struggle when attributes are incomplete or noisy because they assume uniform reliability across all data modalities and nodes. RHEA addresses this by recognizing that the reliability of different data modalities can vary significantly for each node. It leverages the principle of attribute homophily, where connected nodes tend to share similar attributes, to estimate node-specific modality reliability from their graph neighborhoods. This reliability signal is then propagated throughout the clustering process: RHEA reconstructs missing or unreliable modalities, adaptively weights modalities during fusion, and employs a topology-aware optimal transport clustering approach that incorporates reliability-aware assignment and neighbor-consensus distillation. The framework also integrates the confidence of reconstructed representations into its optimization objective, ensuring that uncertain reconstructions contribute appropriately. Extensive experiments demonstrate RHEA's superior performance, especially as attribute quality degrades.