Purely Contrastive Multiview Subspace Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiview subspace clustering (MVSC) integrates information from multiple data views to uncover underlying structures.
• Existing MVSC methods often prioritize sample aggregation within subspaces, neglecting inter-subspace separation.
• This limitation hinders accurate subspace structure identification in complex datasets.

Purpose of the Study:

• To develop a novel MVSC framework that incorporates contrastive learning for improved subspace discovery.
• To enhance the separation of samples across different subspaces, complementing existing aggregation techniques.
• To create a robust method for uncovering the intrinsic subspace structure of multiview data.

Main Methods:

• Introduced a purely contrastive MVSC (PCMVSC) approach by integrating contrastive learning into the MVSC framework.
• Developed a contrastive data self-representation module for enhanced feature learning.
• Incorporated a contrastive regularizer for reconstruction coefficients and a contrastive alignment term for a consensus matrix.

Main Results:

• The proposed modules in PCMVSC demonstrate superiority over similar components in existing methods.
• The consensus reconstruction coefficient matrix effectively reveals the underlying subspace structure of multiview datasets.
• Extensive experiments confirm PCMVSC's effectiveness and its outperformance of various existing multiview clustering algorithms.

Conclusions:

• PCMVSC offers a significant advancement in multiview subspace clustering by leveraging contrastive learning.
• The method effectively addresses the limitations of traditional MVSC by emphasizing both sample aggregation and separation.
• PCMVSC provides a powerful and effective solution for uncovering subspace structures in complex multiview datasets.