Homophily-aware multi-view graph clustering via multi-order filtering

Multi-view graph clustering (MGC) aims to uncover latent semantic structures by integrating complementary information from multiple graph views. However, real-world multi-view graphs often exhibit noisy and inconsistent topologies, along with fixed-order neighborhood propagation that fails to capture structural patterns across diverse homophily levels. These limitations highlight a fundamental challenge of how to construct a unified and homophily-enhanced structural representation that supports scale-aware modeling under heterogeneous conditions. To address this, we propose Homophily-aware Multi-view Graph Clustering via Multi-order Filtering (HMGC-MF), a unified framework that progressively integrates structure refinement and multi-order frequency-aware propagation in a mutually reinforcing manner. Specifically, we construct a semantic-structural consensus graph via edge-wise fusion of feature similarity and multi-view topology, yielding a refined structure that suppresses inter-class noise and strengthens intra-class connectivity. Upon this foundation, we perform multi-order dual-pass filtering to extract low-frequency signals that reflect homophilous patterns and high-frequency components that capture heterophilous ones, adaptively balanced by a learned homophily ratio. To fully exploit cross-view complementarities, we further employ a similarity-based dynamic view fusion strategy and a self-supervised clustering objective to guide representation learning. Extensive experiments on public benchmarks demonstrate that HMGC-MF consistently outperforms state-of-the-art baselines, especially in scenarios with structural noise or weak homophily.