Two-Stage Regularized Linear Discriminant Analysis for 2-D Data.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• Fisher's linear discriminant analysis (LDA) relies on within-class and between-class covariance matrices.
• Regularized LDA (RLDA) improves LDA for 2-D data by regularizing eigenvalues of the within-class matrix.
• Existing methods like RLDA do not fully optimize eigenvectors or the between-class matrix for 2-D data.

Purpose of the Study:

• To propose a novel two-stage method for improving LDA on 2-D data by simultaneously enhancing both within-class and between-class matrices.
• To introduce a bidirectional LDA (BLDA) stage that incorporates row and column correlations inherent in 2-D data.
• To develop a statistical test for determining optimal subspace dimensionality in the initial stage.

Main Methods:

• A two-stage approach combining bidirectional LDA (BLDA) and regularized LDA (RLDA).
• BLDA utilizes separable covariance constraints to address 2-D data correlations.
• A statistical test is employed for subspace dimensionality selection in the BLDA stage.
• RLDA is applied in a reduced-dimensional subspace determined by BLDA.

Main Results:

• The BLDA stage significantly reduces dimensionality while preserving crucial discriminant information.
• The combined BLDA+RLDA method improves both the within-class and between-class matrices.
• Experiments demonstrate superior performance of BLDA+RLDA over competing methods on synthetic and real-world 2-D datasets.

Conclusions:

• The proposed BLDA+RLDA method offers a significant advancement for discriminant analysis in 2-D data.
• Simultaneous improvement of covariance matrices and effective dimensionality reduction are key benefits.
• The method shows robust performance across various 2-D data types.