Multilevel Contrastive Multiview Clustering With Dual Self-Supervised Learning.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Multiview clustering (MVC) integrates diverse data views for improved performance.
• Contrastive learning excels in unsupervised representation learning but has limitations in MVC.
• Existing contrastive MVC methods overlook nearest neighbors and multilevel data structures.

Purpose of the Study:

• To propose a novel end-to-end deep MVC method, Multilevel Contrastive MVC (MCMC).
• To address limitations in existing contrastive MVC by incorporating nearest neighbors and multilevel structures.
• To enhance the compactness and accuracy of multiview clustering through dual self-supervised learning (DSL).

Main Methods:

• Developed MCMC utilizing nearest neighbors from latent subspace as positive pairs for instance-level compactness.
• Implemented multilevel contrastive learning (MCL) on clusters, instances, and prototypes to capture data structure.
• Employed DSL to learn consistent cluster assignments by associating different structural representations.

Main Results:

• MCMC demonstrated improved intracluster compactness and intercluster separability.
• The proposed method achieved higher accuracy (ACC) in clustering performance compared to existing methods.
• The approach effectively captures the multilevel representational structure inherent in multiview datasets.

Conclusions:

• MCMC offers a significant advancement in multiview clustering by leveraging nearest neighbors and multilevel contrastive learning.
• The integration of DSL ensures consistent and accurate cluster assignments across different representation levels.
• The proposed method provides a robust framework for enhancing the performance of unsupervised multiview clustering.