Abstract
Multi-view clustering has achieved promising performance with the advancement in deep neural networks. However, most existing multi-view clustering methods assume that cross-view correspondences are fully known, which is often unrealistic in real-world scenarios due to factors such as temporal asynchrony and sensor faults. To address this, we propose a novel Partial Contrastive Learning (PCL) framework that tackles the practical yet understudied problem of partially aligned multi-view clustering. The PCL combines explicit cross-view correspondence modeling with the contrastive learning paradigm. Specifically, PCL integrates a partial alignment module to mitigate inherent misalignment by adaptively computing soft matching probabilities. These probabilities, in turn, guide a generalized contrastive loss to learn robust representations from both aligned and unaligned data. Overall, our PCL effectively enhances the discriminative power of the learned multi-view representations, achieving superior alignment and clustering performance. We demonstrate the effectiveness of the PCL through extensive comparative experiments with eleven partially aligned multi-view clustering methods on eight benchmark datasets.