Sparse discriminant PCA based on contrastive learning and class-specificity distribution.

Area of Science:

• Machine Learning
• Data Science
• Dimensionality Reduction

Background:

• Principal Component Analysis (PCA) is a widely used feature extraction technique.
• Existing robust PCA methods often focus on reconstruction error and overlook class-specific data distributions, limiting their discriminative power.

Purpose of the Study:

• To propose a Sparse Discriminant Principal Components Analysis (SDPCA) model.
• To address the limitations of traditional PCA by incorporating contrastive learning and class-specificity distribution.

Main Methods:

• SDPCA utilizes contrastive learning to capture discriminative information between samples and their reconstructions.
• Minimizes the squared ℓ1,2-norm of the low-dimensional embedding to reflect class-specificity.
• Applies sparsity constraints via the squared ℓ1,2-norm on the projection matrix for noise reduction and interpretability.

Main Results:

• Experimental results on benchmark databases show state-of-the-art performance.
• The proposed SDPCA model effectively suppresses noise and enhances feature extraction.

Conclusions:

• SDPCA offers a more discriminative and robust approach to Principal Component Analysis.
• The model's ability to leverage class-specific information and sparsity improves feature extraction performance.