A Wrapper Feature Subset Selection Method Based on Randomized Search and Multilayer Structure.

Area of Science:

• Biomedical science
• Machine learning
• Computational biology

Background:

• Clinical diagnosis relies on identifying discriminative features from complex data.
• Machine learning methods have shown promise but often overlook intricate feature interactions.
• Cancer and other diseases result from multifactorial feature interactions.

Purpose of the Study:

• To propose a multilayer feature subset selection method (MLFSSM) for identifying discriminative feature subsets.
• To address limitations of traditional methods that focus on single features.
• To improve the accuracy of clinical diagnosis by considering complex feature interactions.

Main Methods:

• MLFSSM employs a randomized search and multilayer structure.
• Feature subsets are generated and evaluated layer-by-layer to assign weights.
• Feature weights are refined iteratively based on subset performance, prioritizing features in high-performing subsets.

Main Results:

• MLFSSM selected more discriminative feature subsets compared to LVW, GAANN, and SVM-RFE.
• The method demonstrated superior classification performance against state-of-the-art techniques like TSP, K-TSP, and RS-DC.
• Experimental validation on five public gene datasets confirmed the efficacy of MLFSSM.

Conclusions:

• MLFSSM effectively captures complex feature interactions crucial for disease, particularly cancer.
• The proposed method offers improved feature subset selection for enhanced clinical diagnosis.
• MLFSSM represents a significant advancement over existing feature selection techniques.