Prototype Contrastive Learning For Multi-View Clustering
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Multi-view clustering (MVC) aims to integrate heterogeneous information from diverse views to reveal intrinsic data partitions. While recent deep MVC methods, particularly those employing contrastive learning like Multi-level Feature Learning , have made significant strides in separating view-private noise from common semantics, they often suffer from a fundamental limitation: reliance on instance-wise discrimination. This paradigm effectively pulls positive pairs together but neglects the underlying semantic structure of the data, often resulting in loose cluster distributions where semantically similar samples remain dispersed. To address this challenge, we propose a novel framework termed Prototype Contrastive Learning For Multi-View Clustering (PCMVC). PCMVC inherits the robust feature disentanglement of a multi-level architecture but introduces a prototypical contrastive mechanism into the high-level feature space. By treating prototypes as latent variables within an Expectation-Maximization (EM) framework, PCMVC iteratively estimates cross-view prototypes and optimizes the network to align instances with their corresponding semantic centroids. This approach explicitly enforces intra-class compactness and interclass separability globally, rather than locally. Extensive experiments on six benchmark datasets demonstrate that PCMVC has superior clustering accuracy and representation robustness.
Fields of Study & Subjects
Advanced Clustering Algorithms Research
Face and Expression Recognition
Machine Learning and Data Classification
Publication details
- Journal / source
- International Conference on Communication Networks and Machine Learning (CNML)
- Pages
- 1296–1299
- Publication date
- 30 January 2026
- Language
- EN