Cross-view alignment and completion for incomplete information multi-view clustering

Multi-view clustering seeks to group data instances by exploiting the complementary and comprehensive information from multiple views. Due to the complexities involved in data acquisition and alignment, practical challenges such as the Partially View-unaligned Problem (PVP) and the Partially Sample-missing Problem (PSP) may emerge, and the multi-view problem with incomplete information when both coexist. Current research primarily focuses on solving PVP or PSP individually. Although there is a method that can handle PVP and PSP simultaneously, it relies on traditional Euclidean distance to compute cross-view distance matrices. This separation of matrix computation and network training hinders achieving globally optimal optimization matrices and fails to capture the consistency and complementarity between views adequately. To overcome these challenges, the Cross-View Alignment and Completion for Incomplete Information Multi-View Clustering (CAC-MVC) is proposed. In brief, CAC-MVC integrates a contrastive optimization module to improve the consistency between feature representation and alignment tasks. In addition, a module based on category-level alignment and completion strategy is designed to establish optimal alignment and completion relationships between views. Our experiments on four datasets, alongside comparisons with over 10 state-of-the-art approaches, demonstrate the effectiveness and robustness of CAC-MVC in multi-view clustering with incomplete information.