Bi-Contrastive Alignment for Incomplete Multi-view Clustering

Multi-view clustering (MVC) represents a significant challenge in the realms of machine learning and data mining. Recently, the advent of contrastive learning (CL) has catalyzed the development of numerous innovative methods within this field. The efficacy of these methods typically hinges on the availability of consistent and comprehensive data. Nevertheless, real-world scenarios of data collection and transmission frequently encounter issues like partial unaligned-views (PU), which result in incomplete datasets and pose formidable challenges for effective MVC. To address this prevalent issue, our paper introduces an innovative framework, termed bi-contrastive alignment for incomplete multi-view clustering (BCA). BCA specifically targets the challenge of incomplete data stemming from PU. It encompasses a unique module, semantic guided bi-contrastive alignment (SGBA), designed to counteract the PU by harnessing semantic similarities. Moreover, the framework augments data alignment by integrating a contrastive learning-based corrective component into the loss function. The comparative experiments demonstrate that the ARI metric of our method surpasses the second-best contrasting method on all four datasets.