Regularized F-measure maximization for feature selection and classification.

Area of Science:

• Biostatistics
• Machine Learning
• Bioinformatics

Background:

• Receiver Operating Characteristic (ROC) analysis is widely used for classification performance assessment, particularly in medical diagnostics and biological marker evaluation.
• Traditional ROC analysis and utility functions like F-measure are valuable when misclassification costs are unknown, common in real-world scenarios.
• F-measure, a combination of precision and recall, offers a global performance metric.

Purpose of the Study:

• To propose a novel method for classification performance assessment using regularized F-measure maximization.
• To address challenges in highly unbalanced datasets and scenarios with missing labels for samples.
• To integrate simultaneous feature selection and prediction within the proposed framework.

Main Methods:

• A novel method based on regularized F-measure maximization is proposed.
• The method incorporates differential costs for positive and negative samples.
• Simultaneous feature selection and prediction are achieved using an L(1) penalty.

Main Results:

• Experimental results on benchmark, methylation, and high-dimensional microarray data are presented.
• The proposed algorithm demonstrates superior or equivalent performance compared to other popular classifiers in limited experiments.
• The method is particularly effective for highly unbalanced datasets and datasets with missing negative or positive sample labels.

Conclusions:

• The proposed regularized F-measure maximization method offers an effective approach for classification, especially in challenging data scenarios.
• The integration of feature selection and prediction enhances its utility for complex biological and medical data.
• This method provides a valuable tool for evaluating diagnostic tests and biological markers.