Threshold Adaptation for Improved Wrapper-Based Evolutionary Feature Selection.

Area of Science:

• Computational intelligence
• Machine learning
• Data science

Background:

• High-dimensional datasets require effective feature selection for improved classification accuracy, reduced overfitting, and enhanced interpretability.
• Evolutionary Feature Selection (EFS) commonly uses a static threshold (θ=0.5) for feature inclusion, which may not be optimal.

Purpose of the Study:

• To conduct the first large-scale, systematic evaluation of threshold adaptation mechanisms in wrapper-based EFS.
• To compare deterministic, adaptive, and self-adaptive threshold control strategies against static thresholds.

Main Methods:

• A unified framework was developed to integrate various threshold parameter control mechanisms within arbitrary bio-inspired algorithms.
• Extensive experiments were performed on diverse benchmark datasets, analyzing classification accuracy, feature subset size, and convergence properties.

Main Results:

• Adaptive threshold mechanisms significantly outperformed static threshold control in EFS.
• Adaptive methods achieved superior trade-offs between classification accuracy and feature subset size.
• The proposed adaptive EFS methods surpassed state-of-the-art feature selection techniques on multiple benchmarks.

Conclusions:

• Threshold adaptation plays a critical role in optimizing Evolutionary Feature Selection.
• Adaptive mechanisms offer practical guidelines for enhancing EFS performance and achieving better results in high-dimensional data analysis.