.png){kind=icon}
.png){kind=icon}
Return
Adaptive deep multi-view clustering via quality-aware representation learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.115619.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep multi-view clustering aims to uncover the complex non-linear correlations embedded within multi-view data by leveraging deep learning. However, existing studies often suffer from two major limitations: sensitivity to low-quality or noisy views, and misalignment between representation learning objectives and clustering task. This paper constructs an adaptive and dynamic multi-view clustering framework. It employs a teacher-student asymmetric learning paradigm to guide shared consistency across views, while leveraging the Variational Information Bottleneck (VIB) principle to regularize student views and preserve their crucial complementary information. Specifically, A quality-aware representation learning strategy is proposed, which leverages teacher-student cross-view guidance and VIB-based regularization to extract complementary features and mitigate noise and redundancy. To simultaneously enhance semantic consistency, extract discriminative feature information, and balance cluster assignments, a composite loss function-integrating cross-view contrastive, VIB, and entropy regularization terms-is further constructed. Furthermore, an adaptive fusion strategy incorporating static, dynamic, and attention-based modes is developed. Cross-view contrastive learning is then applied to boost sample assignment consistency across views, yielding an adaptive deep multi-view clustering algorithm. Extensive experiments on multiple benchmark datasets show that the proposed method not only enhances cross-view consistency and information integration but also achieves superior performance and robustness across delivers multi-view clustering scenarios.
Keywords:
Deep multi-view clustering
Adaptive framework
Quality-aware representation learning
Variational Information Bottleneck
Cross-view contrastive learning
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W