CRRC: Residual Cross-view Learning for Deep Multi-view Clustering

Xiaojian Ding, Xiaoying Zhu, Lin Zhao, Xian Li

Keywords: Deep Multi-View Clustering, Contrastive Learning, Residual Connections, Dynamic Gating, Attention-Based Weighting

TL;DR: CRRC

Abstract: Deep multi-view clustering aims to integrate complementary information from multiple heterogeneous views to improve clustering performance. However, existing fusion strategies often struggle to balance shared semantics and view-specific heterogeneity, as they typically rely on direct concatenation or rigid alignment, which obscures subtle cross-view patterns and assumes equal contribution from all views. To address these limitations, we propose CRRC, a novel framework that leverages residual connections to recalibrate view-specific features by adaptively incorporating complementary information from other views. Specifically, CRRC introduces a dynamic gating fusion module to control residual flow based on view characteristics, and an attention-based weighting mechanism to emphasize semantically relevant cross-view signals. These components work collaboratively to enhance feature discriminability and consistency. Extensive experiments on benchmarks datasets demonstrate that CRRC outperforms state-of-the-art methods in accuracy, NMI, and purity, validating its effectiveness in achieving robust multi-view clustering.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 3235