Pairwise constraint weighted imputation for incomplete multi-view clustering with high missing rate

Incomplete multi-view clustering (IMVC) has garnered considerable attention for its capacity to manage real-world scenarios involving missing multi-view data. However, existing deep IMVC approaches typically exhibit two key limitations: 1) they inadequately exploit limited supervisory information to enhance the clustering performance, and 2) their effectiveness deteriorates substantially under high missing rates due to severe information sparsity. To overcome these challenges, we propose a novel deep imputation-based semi-supervised IMVC framework, termed pairwise constraint weighted imputation (PCWI), specifically designed for incomplete multi-view clustering in high missing rate scenarios. In this framework, available pairwise constraint supervisory information is employed to identify the most appropriate samples for high-quality reconstruction of the missing data. Moreover, to maximize the utility of a small amount of supervisory information, we incorporate adaptive contrastive learning to refine inter-view similarities at the view level. Simultaneously, pairwise constraint information is leveraged at the sample level to encourage intra-class compactness and inter-class separability. To the best of our knowledge, this is the first study to leverage pairwise constraint supervisory information for view imputation in IMVC tasks. Extensive experiments conducted on six multi-view datasets under high missing rates demonstrate that the proposed PCWI method outperforms several state-of-the-art methods, validating its effectiveness.