Novel Regularization Method for Biomarker Selection and Cancer Classification.

Area of Science:

• Machine Learning
• High-Dimensional Data Analysis
• Bioinformatics

Background:

• Variable selection is crucial in big data and machine learning, especially for high-dimensional datasets like gene expression data.
• Regularization methods, using L2 and fixed q penalties, are common but often require specific data distributions.
• Identifying relevant variable groups, such as biological pathways, is a key challenge.

Purpose of the Study:

• To propose a novel complex harmonic regularization (CHR) penalty function for effective group variable selection.
• To develop a method that approximates combined L_p and L_q regularizations with adjustable parameters.
• To enhance variable selection and classification performance in high-dimensional data analysis.

Main Methods:

• Introduction of a new complex harmonic regularization (CHR) penalty function.
• Development of a direct path seeking algorithm for solving the CHR penalty.
• Adjustable parameters p and q to control sparsity and grouping effects.

Main Results:

• The proposed CHR penalty function effectively selects groups of relevant variables.
• CHR demonstrates superior performance compared to existing state-of-the-art regularization methods.
• Improved classification accuracy using the CHR method.

Conclusions:

• Complex harmonic regularization (CHR) provides a flexible and powerful approach for group variable selection.
• CHR overcomes limitations of existing methods by not requiring specific data distributions.
• The CHR method offers significant advantages in both variable selection and classification tasks.