Contrastive-Driven Diversity and Consistency Exploration in Tensorized Multiview Subspace Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiview subspace clustering (MVSC) integrates information from multiple data views.
• Existing MVSC methods often fail to leverage correlations between consensus and complementary information.
• There is a need for MVSC algorithms that can model both positive and negative correlations within data representations.

Purpose of the Study:

• To propose a novel MVSC method, contrastive-driven diversity and consistency exploration in tensorized MVSC (CD-TMSC).
• To effectively integrate consensus and complementary information by modeling both positive and negative correlations.
• To improve clustering performance by enhancing representation diversity and consistency.

Main Methods:

• Segmenting self-representations into consensus and specific representations.
• Introducing a novel fractional regularization term inspired by contrastive learning, utilizing the Hilbert-Schmidt independence criterion (HSIC).
• Incorporating graph regularization for consensus matrices and low-rank tensor constraints for higher-order correlations.

Main Results:

• The proposed CD-TMSC method effectively models both consensus and complementary information.
• The contrastive-driven regularization amplifies negative correlations, promoting diversity, and reinforces positive correlations, enhancing consistency.
• Experimental results demonstrate superior performance compared to state-of-the-art MVSC methods on benchmark datasets.

Conclusions:

• CD-TMSC offers a cohesive framework integrating contrastive learning, manifold learning, and tensor learning for MVSC.
• The method successfully addresses the limitations of existing MVSC algorithms by exploiting inter-representation correlations.
• The proposed approach achieves state-of-the-art performance, highlighting the effectiveness of the integrated strategy.