Sparse Multicategory Generalized Distance Weighted Discrimination in Ultra-High Dimensions.

Area of Science:

• Machine Learning
• Statistical Classification
• Bioinformatics

Background:

• Distance Weighted Discrimination (DWD) effectively handles data piling in high dimensions.
• Variable selection in multicategory classification with sparse structures presents significant challenges.

Purpose of the Study:

• To propose a novel multicategory generalized DWD (MgDWD) method.
• To incorporate a sparse group lasso penalty for maintaining intrinsic variable group structures during selection.

Main Methods:

• Developed the multicategory generalized DWD (MgDWD) method.
• Utilized a sparse group lasso penalty for variable selection.
• Derived theoretical properties including minimizer uniqueness and classifier consistency.
• Implemented an efficient proximal operator-based algorithm for optimization.

Main Results:

• The proposed MgDWD method maintains variable group structures.
• Theoretical guarantees for minimizer uniqueness and classifier consistency were established.
• The algorithm efficiently solves the optimization problem.

Conclusions:

• MgDWD offers a robust solution for variable selection in multicategory classification with high-dimensional data.
• The method demonstrated effective performance in simulations and a real-world miRNA HIV study.