Prototype Matching Learning for Incomplete Multi-View Clustering

Abstract

As information acquisition diversifies, data is acquired and stored in increasing modalities. However, sensor failures or equipment issues can lead to partial data loss in certain views, resulting in incomplete multi-view clustering (IMVC) problems. Although some prototype-based IMVC methods have achieved satisfactory performance, almost all of these methods implicitly assume that the cross-view prototypes are aligned. However, during the generation or selection of prototypes, different networks could produce different prototypes, thereby leading to potential misalignment of prototypes across views, i.e., prototype-unaligned problem (PUP). The presence of PUP could lead to overfitting the model. Additionally, when recovering the missing data, there is uncertainty in data quality under different missing rates, which could lead to the performance instability problem (PIP). To address these issues, we propose Prototype Matching Learning for Incomplete Multi-view Clustering (PMIMC). Specifically, PMIMC leverages relational consistency learning to mitigate the heterogeneity of multi-view data. Subsequently, we design a robust prototype contrastive learning loss for the generated prototypes to reduce the effects of PUP. Finally, we propose a prototype-based imputation strategy, that aims to alleviate the instability of imputation under high missing rates. Extensive experiments demonstrate that PMIMC outperforms 13 state-of-the-art methods in terms of clustering performance and robustness. The code is available at: https://github.com/hl-yuan/PMIMC.