Multi-view Subspace Classification: A Hierarchical Contrastive Approach and Low-rank Latent Representation

Abstract

Effective multi-view subspace learning is crucial for enhancing classification performance on multi-view data. In this paper, we propose CMvLSCN, a novel end-to-end framework addressing multi-view classification at view, sample, and subspace levels. The key innovations are: Strengthening inter-view consistency within categories while weakening inter-view similarity across categories; Learning a unified latent subspace representation through the view fusion; and Imposing low-rank latent self-representation and hierarchical contrastive constraints to better classify multi-view data. CMvLSCN employs contrastive learning to optimize Kullback-Leibler divergence among views, imposes low-rank structure on the latent subspace, and introduces sample-level contrastive constraints. This approach captures underlying data relationships and enhances subspace representation discriminability. Experiments demonstrate superior performance, especially with limited training data. Code and datasets are available on GitHub.