Robust Hierarchical Contrastive Learning for Multi-view Clustering with Imbalanced Views

Contrastive learning has become a dominant paradigm for multi-view clustering due to its ability to align representations across views while preserving inter-sample discriminability. However, existing contrastive multi-view clustering methods generally treat all views as equally reliable and enforce uniform cross-view alignment. In real-world applications, where views often exhibit heterogeneous distributions and imbalanced quality, such forced alignment can undermine representation learning by propagating noise from low-quality views to high-quality ones. To address this challenge, we propose a novel framework called Robust Hierarchical Contrastive Learning for Multi-view Clustering with Imbalanced Views (RCHMC-IV). Specifically, RCHMC-IV uses a weighted cross-view learning mechanism that explicitly accounts for both feature-level discrepancies and cluster-level distribution shifts across views. By adaptively weighting features and clusters, RCHMC-IV facilitates discriminative cross-view representation learning and mitigates the risk of degrading high-quality views through forced alignment with inferior ones. Extensive experiments across twelve datasets demonstrate that RCHMC-IV significantly outperforms existing state-of-the-art approaches.