Robust Semantic Sample Filtering for Partially View-aligned Clustering

Yiyou Gao, Zeren Sun, Yazhou Yao, Zechao Li

Keywords: Multi-view Clustering; Partially View-aligned Clustering

Abstract:

Multi-view Clustering (MvC) typically assumes strict sample alignment across views.
However, this assumption often fails to hold in real-world scenarios due to data acquisition or occlusions, resulting in partially view-aligned data.
Existing methods tend to rely on prior alignment knowledge and discard unaligned samples during training, hindering their performance and practical applicability.
To address this, we propose a novel framework named REFINE that integrates Cross-view Semantics-based Filtering and Shared-space Contrastive Learning to robustly handle partially view-aligned data.
Our method dynamically identifies reliable samples by aligning pseudo-labels across views and filters out noisy correspondences to improve clustering prototype initialization and cross-view consistency learning.
Moreover, we employ a cross-view decoder to project features into a shared latent space, bridging modality gaps and facilitating more effective contrastive learning.
Extensive experiments across five benchmark datasets under both fully aligned and partially aligned settings demonstrate that our approach achieves state-of-the-art performance, delivering superior robustness and generalization in real-world scenarios without strict alignment requirements.
Our code has been made anonymously available at https://github.com/REFINE-REFINE/REFINE.

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

Submission Number: 10177