Abstract
Feature selection is essential for enhancing classification accuracy, reducing overfitting, and improving interpretability in high-dimensional datasets. Evolutionary Feature Selection (EFS) methods employ a threshold parameter θ to decide feature inclusion, yet the widely used static setting θ=0.5 may not yield optimal results. This paper presents the first large-scale, systematic evaluation of threshold adaptation mechanisms in wrapper-based EFS across a diverse number of benchmark datasets. We examine deterministic, adaptive, and self-adaptive threshold parameter control under a unified framework, which can be used in an arbitrary bio-inspired algorithm. Extensive experiments and statistical analyses of classification accuracy, feature subset size, and convergence properties demonstrate that adaptive mechanisms outperform the static threshold parameter control significantly. In particular, they not only provide superior tradeoffs between accuracy and subset size but also surpass the state-of-the-art feature selection methods on multiple benchmarks. Our findings highlight the critical role of threshold adaptation in EFS and establish practical guidelines for its effective application.