Skip to content
Book Open access

Resilient Deep Multi-View Clustering via Dual-Time-Scale Kernel Attention

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 30 references

Abstract

Multi-view clustering aims to cluster data into different groups by leveraging both the consistency and complementarity across multiple views. Nevertheless, existing methods commonly face two primary limitations. First, cross-view information fusion often lacks fine-grained sample-level modeling, leading to semantic drift during feature alignment. Second, the dynamic optimization process suffers from inherent inertia, failing to correct accumulated errors stemming from infrequent pseudo-label updating during training. To address these problems, this paper proposes a novel framework named Resilient Deep Multi-View Clustering via Dual-Time-Scale Kernel Attention (DKAMVC). Specifically, we first adopt an autoencoder coupled with a capacity-adaptive MLP to extract latent representations. Then, the unified features obtained by connecting the latent representations are fed into a dual-time-scale model which integrates slow anchor updating with fast weight learning for resilient noise suppression. Next, these reinforced features are directed into a decoupled dual-pathway architecture, which utilizes a Local Kernel Attention (LKA) module to accurately quantify localized confidence. This design completely separates distribution modeling and weighting functions to establish fine-grained sample-to-target matching in the Reproducing Kernel Hilbert Space (RKHS). Finally, the multiply-augmented global features are clustered to yield the ultimate results. Experiments conducted on eight multi-view data sets demonstrate that the clustering performance of DKAMVC outperforms state-of-the-art comparison methods.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.