Reproduce “Cross-Contrastive Clustering for Multimodal Attributed Graphs with Dual Graph Filtering” on OpenResearch

About this paper

Haoran Zheng, Renchi Yang, Hongtao Wang, Jianliang Xu

The Dual Graph Filtering (DGF) framework provides a scalable approach for clustering multimodal attributed graphs by employing dual-domain graph filtering for robust representation learning and a tri-cross contrastive training strategy. It consistently outperforms 14 baselines across eight real-world datasets, achieving, for instance, a +10.53% accuracy gain and +10.35% adjusted Rand index increase on Reddit-S, and successfully processes large graphs where other methods fail.

Key insights

• Effective MMAG clustering requires simultaneous low-pass filtering in both the node domain (graph topology) and the feature domain (attribute dimensions) to generate robust and discriminative representations.
• Contrastive learning strategies must be tailored for MMAGs by explicitly addressing low inter-modality correlation, refining graph neighborhoods through attribute-aware pruning, and strengthening community structures via hard positive sampling.
• Theoretical analysis confirms that DGF acts as a low-pass filter in both domains, preserving essential smooth structures while attenuating high-frequency noise in both the graph and its features.

Summary by alphaXiv.