Robust Discriminant Regression for Feature Extraction.

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• Ridge Regression (RR) and its variants are common for feature extraction but suffer from data sensitivity and limited projection capabilities.
• Existing RR methods struggle with variations in data, impacting feature extraction and recognition performance.

Purpose of the Study:

• To develop a robust feature extraction method that addresses the limitations of traditional Ridge Regression.
• Introduce Robust Discriminant Regression (RDR) to enhance feature extraction robustness and accuracy in pattern recognition.

Main Methods:

• Propose Robust Discriminant Regression (RDR) employing the L2,1-norm for enhanced robustness.
• Develop an iterative algorithm involving an eigenfunction to solve the robust objective function.
• Utilize eigen decomposition to obtain optimal orthogonal projections within the RDR framework.

Main Results:

• RDR demonstrates superior performance compared to classical and recent regression-based methods.
• Experiments confirm RDR's effectiveness across various databases, outperforming both L2-norm and L2,1-norm regression techniques.
• Convergence analysis and computational complexity of the RDR algorithm are presented.

Conclusions:

• Robust Discriminant Regression (RDR) offers a significant advancement in feature extraction for pattern recognition.
• The proposed L2,1-norm-based RDR method provides enhanced robustness and accuracy over existing techniques.
• The study validates RDR's effectiveness through empirical experiments on benchmark datasets.