Graph-based sparse linear discriminant analysis for high-dimensional classification.

Area of Science:

• Machine Learning
• Statistical Classification
• Data Mining

Background:

• Linear Discriminant Analysis (LDA) is a standard classification method, but struggles with high-dimensional data.
• Existing high-dimensional LDA variants often fail to leverage predictor structure information.
• There is a need for advanced LDA techniques that incorporate feature relationships for improved performance.

Purpose of the Study:

• To introduce a novel high-dimensional LDA method, Graph-based Sparse LDA (GSLDA), that utilizes feature graph structure.
• To develop a semi-supervised variant of GSLDA for leveraging unlabeled data.
• To establish theoretical properties and demonstrate practical performance of the proposed methods.

Main Methods:

• Proposed GSLDA uses a regularized regression formulation with a structure-based sparse penalty on the discriminant vector.
• The feature graph structure can be predefined or learned from training data.
• A semi-supervised GSLDA variant is developed by exploring the relationship between within-class and overall feature structures.

Main Results:

• GSLDA yields sparse estimates of the discriminant vector, leading to more accurate and interpretable classifiers.
• Theoretical guarantees for selection consistency and convergence rate of the classifier are established.
• The proposed GSLDA demonstrates competitive performance on both simulated and real-world datasets.

Conclusions:

• GSLDA offers a powerful approach for high-dimensional classification by integrating feature structure information.
• The semi-supervised variant effectively utilizes abundant unlabeled data.
• GSLDA provides a robust and interpretable alternative to existing LDA methods for complex datasets.