Inter-class sparsity based discriminative least square regression.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Least square regression is a popular supervised classification method.
• Existing methods overlook sample correlations and use inappropriate zero-one label matrices, limiting performance.
• Addressing these limitations is crucial for advancing multi-class classification.

Purpose of the Study:

• To propose a novel method, inter-class sparsity based discriminative least square regression (ICS_DLSR), for multi-class classification.
• To enhance classification performance by addressing limitations of traditional least square regression.
• To develop a more discriminative transformation matrix for improved regression and classification.

Main Methods:

• Introduced inter-class sparsity constraint to reduce intra-class margins and enlarge inter-class margins.
• Incorporated a row-sparsity constrained error term to relax the strict zero-one label matrix.
• Developed a discriminative least square regression model focusing on common sparsity structures within classes.

Main Results:

• The proposed ICS_DLSR method demonstrated superior performance in multi-class classification tasks.
• Experimental results confirmed the effectiveness of the inter-class sparsity and row-sparsity constraints.
• The method learned a more compact and discriminative transformation matrix compared to other approaches.

Conclusions:

• ICS_DLSR effectively addresses the limitations of traditional least square regression for multi-class classification.
• The novel constraints lead to improved discriminative power and classification accuracy.
• The proposed method offers a promising advancement in supervised classification techniques.