Hypergraph-Based Unaligned Multi-View Clustering via Cluster-Aware Feature Extraction

The Unaligned Multi-view Clustering (UMC) aims to mine distinct cluster structures from unaligned multi-view data where sample features are not aligned across different views. Most existing methods extract low-dimensional representations from high-dimensional raw features, preserving key components for alignment and fusion. However, these feature extraction strategies often neglect the sample cluster structure, yielding the extracted representations with ambiguous cluster distributions that lead to subsequent suboptimal alignment and fusion. To address this issue, we propose a hypergraph-based cluster-aware alignment clustering framework (HG-UMC), which provides cluster-aware hyperedges for the hypergraph neural network, mining cluster-aware representations for subsequent effective alignment and fusion. Specifically, we design a Hyperedge-driven Feature Extraction module that constructs multiple groups of learnable hyperedges to capture the intrinsic clustering structure of samples precisely. These cluster-aware hyperedges are incorporated into the feature extraction process of the hypergraph neural network, enabling it to learn cluster-aware representations. In the subsequent alignment phase, we design a Hyperedge-driven Cluster-level Alignment and Fusion module. This module extracts the similarity structure among intra-cluster hyperedges to serve as the cluster-specific topology for cross-view cluster alignment, and incorporates a cluster-wise cross-attention mechanism to achieve cross-view cluster-level fusion. Extensive experiments on various datasets have verified the superiority of HG-UMC over other state-of-the-art methods.