Sparse tensor discriminant analysis.

Area of Science:

• Machine Learning
• Computer Vision
• Pattern Recognition

Background:

• Linear discriminant analysis (LDA) is a foundational technique in pattern recognition.
• Recent advancements have extended LDA to multilinear cases, but sparse extensions are less explored.
• Feature extraction methods are crucial for improving classification accuracy, particularly in high-dimensional data.

Purpose of the Study:

• To introduce a novel sparse discriminant subspace learning method called sparse tensor discriminant analysis (STDA).
• To extend existing multilinear discriminant analysis techniques to incorporate sparsity for enhanced feature extraction.
• To develop an iterative optimization approach for learning sparse discriminant subspaces.

Main Methods:

• The proposed STDA method incorporates L1 and L2 norms into its objective function.
• It utilizes k-mode optimization and L1 norm sparse regression for iterative subspace learning.
• The method identifies important variables/factors by selecting non-zero elements within subspaces.

Main Results:

• STDA successfully generates multiple interrelated sparse discriminant subspaces.
• Extensive experiments on face and action recognition databases demonstrate STDA's effectiveness.
• The algorithm achieves competitive performance compared to existing tensor-based methods, outperforming them in small sample size scenarios.

Conclusions:

• Sparse tensor discriminant analysis (STDA) offers a powerful new approach for feature extraction.
• The method's ability to create sparse, interrelated subspaces enhances its potential for pattern recognition.
• STDA demonstrates significant advantages over other discriminant subspace methods, particularly when dealing with limited training data.