Relative Comparison-Based Consensus Learning for Multi-View Subspace Clustering

Current multi-view subspace clustering methods typically consist of a within-view module, which explores inherent characteristics using the self-expressive coefficient matrix, and a cross-view module, which promotes consensus among all views toward similar strengths. However, the self-expressive coefficients are directly influenced by the characteristics and distributions of input features, and coefficient matrices with varying strengths may indicate the same clustering structure. Therefore, directly regularizing the coefficient matrices towards a common matrix is unnecessary and may even diminish the clustering performance. We find that it is the relative data relationship, rather than the absolute similarity, that plays a pivotal role in clustering. Building on this realization, we propose a relative comparison measure that enables a more contextual understanding of the data relationship. Subsequently, we develop a Relative Comparison-based Consensus Learning (RCCL) model for multi-view subspace clustering, which encourages the relative data similarities to be consistent across different views. Our RCCL model advances in identifying the underlying data relationship, avoiding unnecessary constraints on absolute consistency, and thereby delving into the fundamental nature of multi-view consensus. We introduce an elegant transformation operator for relative comparison and solve RCCL under the framework of alternating direction method of multipliers. Extensive experiments unequivocally demonstrated the superiority of RCCL.