Norm Discriminant Eigenspace Transform for Pattern Classification.

Area of Science:

• Machine Learning
• Data Science
• Pattern Recognition

Background:

• Traditional supervised dimensionality reduction (DR) methods often assume unimodal Gaussian distributions.
• These methods define interclass scatter based on class means, limiting their applicability to complex data.

Purpose of the Study:

• To introduce a novel DR approach, Norm Discriminant Eigenspace Transform (NDET), that utilizes average norms for interclass separability.
• To develop a method robust to multimodal and non-Gaussian data distributions.
• To create a nonlinear version (kernel NDET) for capturing complex feature relationships.

Main Methods:

• NDET characterizes interclass separability using average norms (L2) of classes and intraclass compactness using within-class distance.
• Derivation of an upper bound for NDET and exploration of its solution space for optimal dimensionality reduction.
• Development of kernel NDET for nonlinear data modeling.

Main Results:

• NDET effectively overcomes limitations of traditional DR algorithms tied to unimodal and specific data distribution assumptions.
• Experimental validation on synthetic data demonstrates NDET's superiority.
• Comparative studies on UCI machine learning repository and face recognition datasets confirm NDET's novelty and effectiveness.

Conclusions:

• NDET offers a robust and flexible dimensionality reduction technique adaptable to diverse data distributions.
• The method enhances class discrimination by focusing on average norms and within-class compactness.
• NDET and its kernelized version represent a significant advancement in supervised dimensionality reduction.