Partially View-Aligned Multi-View Clustering based on Progressive Global Consensus Learning

Due to spatiotemporal asynchrony in practical environments, instances may remain unaligned across multi-view data, restricting the exploration of cross-view consensus. Al-though existing methods have achieved notable improvements, global correlations consensus learning (GCL), a family of methods proven effective in fully aligned multi-view clustering, remains largely unexplored in partially view-aligned settings. To address this issue, we propose a novel Partially View-aligned multi-view Clustering based on Progressive Global Consensus Learning (PVC-PGCL), which explores and leverages cross-view consensus from global correlations for promising clustering results. Specifically, Unlike conventional one-step GCL formulations that directly enforce global consistency on raw features, PVC-PGCL adopts a progressive consensus learning strategy that transforms the global correlations consensus learning goal into a phased and interactive learning process. It introduces three types of matrices: primary coefficient matrices, buffer matrices, and global coefficient matrices. Collectively, these matrices form an information transfer layer connecting partially view-aligned data with global correlations; individually, each serves distinct functions such as representation dimensionality reduction, sample reconstruction, global correlation modeling, and consensus learning. Through this layered design, PVC-PGCL achieves multi-level consensus exploration from local to global in data misalignment scenarios. Clustering experiments on several popular multi-view datasets demonstrate the effectiveness and competitiveness of our method.