Cross-view discrepancy-driven dynamic weighting for missing view completion in incomplete multi-view clustering.

Abstract

Incomplete Multi-view Clustering (IMVC) aims to uncover the underlying shared clustering structure across different views in the presence of missing information from the views. While numerous view recovery strategies have been proposed to tackle this prominent and challenging problem, most overlook the noise introduced during data recovery, which inevitably degrades clustering performance. To mitigate this issue, we propose a novel dynamically weighted view completion method that leverages cross-view discrepancy information to enhance both view recovery quality and clustering performance. Specifically, we employ cross-view contrastive learning to learn cross-view consistency, which indirectly measures cross-view discrepancies. Since the noise introduced during view recovery is a primary source of cross-view discrepancies in the imputed data, we utilize the learned consistency features to construct a weight matrix that evaluates the quality of the recovered data. To further suppress external noise, both the imputed data and the weight matrix are fed back into the view completion process, refining the recovered views through an instance-level weighted view recovery loss. Additionally, by iteratively optimizing missing view completion and discrepancy learning, our dynamic weighting strategy progressively reduces noise and enhances clustering performance. Extensive experiments on multiple incomplete benchmark datasets demonstrate that our method outperforms state-of-the-art approaches in both missing view completion and clustering performance. The code is available at the following repository: https://anonymous.4open.science/r/DWMVC-2E5C.