ClearF++: Improved Supervised Feature Scoring Using Feature Clustering in Class-Wise Embedding and Reconstruction.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning for healthcare

Background:

• Feature selection is crucial for disease classification and biomarker identification.
• Information-theoretic methods are common but computationally expensive.
• Previous ClearF method improved efficiency but had unstable feature selection due to bottleneck layer issues.

Purpose of the Study:

• To introduce ClearF++, an improved feature selection method addressing ClearF's limitations.
• To enhance biomarker detection through simplified bottleneck selection and feature-wise clustering.
• To evaluate ClearF++'s performance against existing methods in prediction accuracy and stability.

Main Methods:

• ClearF++ simplifies bottleneck layer selection compared to ClearF.
• Incorporates feature-wise clustering, utilizing the Deep Embedded Clustering (DEC) algorithm.
• Performance compared against MultiSURF, IFS, and ClearF on benchmark datasets.

Main Results:

• ClearF++ demonstrates superior prediction accuracy and stability over other methods, particularly with limited samples.
• Deep Embedded Clustering integration enhances performance, showing suitability for complex, small-sample datasets.
• ClearF++ offers faster execution times.

Conclusions:

• ClearF++ provides a more stable and accurate feature selection approach for disease classification and biomarker discovery.
• The method is highly effective and valuable for biomedical data analysis, especially when dealing with limited sample sizes.
• Simplified bottleneck selection and DEC-based clustering are key improvements for biomarker prioritization.