Selective Contrastive Learning for Unpaired Multi-View Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Unpaired multi-view clustering (UMC) is an emerging challenge where paired samples across views are unavailable.
• Existing incomplete multi-view clustering methods rely on sample pairing, which is invalid for UMC.
• Effective joint clustering requires leveraging unpaired samples across multiple views.

Purpose of the Study:

• To address the novel issue of unpaired multi-view clustering (UMC).
• To propose a method that mines consistent cluster structures between views without paired data.
• To overcome challenges of uncertain clustering structures and uncertain cluster pairing in UMC.

Main Methods:

• Developed selective contrastive learning for UMC (scl-UMC).
• Introduced an inner-view (IV) selective contrastive learning module to enhance clustering structures by selecting confident samples.
• Designed a cross-view (CV) selective contrastive learning module for iterative cluster matching and tightening, enhanced by mutual information.

Main Results:

• The proposed scl-UMC method demonstrates significant efficiency in unpaired multi-view clustering.
• Experimental results show superior performance compared to existing state-of-the-art methods.
• The method successfully addresses the challenges of unsupervised clustering structure uncertainty and uncertain cluster pairing.

Conclusions:

• Selective contrastive learning provides an effective solution for unpaired multi-view clustering.
• The proposed inner-view and cross-view contrastive learning modules enhance clustering accuracy and robustness.
• This work advances the field of multi-view clustering by enabling joint clustering with unpaired data.