Deep Subspace Clustering Under Class Relation Constraint.

Area of Science:

• Machine Learning
• Computer Vision
• Data Mining

Background:

• Deep subspace clustering (DSC) relies on latent features for self-expression coefficient matrix construction.
• Current DSC methods often overlook the critical role of latent feature quality.
• Improved latent features are essential for accurate self-representation and effective clustering.

Purpose of the Study:

• To propose a novel Class Relation Constraint (CRC) induced Deep Subspace Clustering (DSC) method.
• To enhance the representation capabilities of latent features within DSC frameworks.
• To improve the accuracy of the self-expression coefficient matrix through better latent feature learning.

Main Methods:

• Developed a Class Relation Constraint (CRC) to optimize latent features in DSC.
• Implemented an intra- and inter-class weighted constraint for enhanced latent data separability.
• Introduced a contrastive loss function within diagonal blocks of the self-expression coefficient matrix, guided by spectral clustering.

Main Results:

• The proposed CRC-DSC method significantly improves the representation ability of latent features.
• Enhanced latent features lead to a more accurate self-expression coefficient matrix.
• Experimental validation on benchmark datasets confirms the method's effectiveness, particularly for small samples and complex data.

Conclusions:

• The CRC-induced DSC method offers a superior approach to subspace clustering by focusing on latent feature enhancement.
• The method demonstrates robust performance on challenging datasets, outperforming existing techniques.
• This work highlights the importance of integrating class relationships for effective deep subspace clustering.