An adaptive length-variation based evolutionary multitasking algorithm for feature selection of high-dimensional classification

Abstract

Evolutionary multitasking (EMT) has recently gained attention as a promising and efficient paradigm for feature selection (FS) in high-dimensional classification problems. However, most existing EMT-based FS approaches rely on fixed-length coding schemes, which force the algorithm to search within the large original feature space. This often leads to increased computational complexity and reduced search efficiency. To overcome these limitations, this paper proposes a novel EMT-based algorithm with an adaptive length variation mechanism, called EMT-ALV. The proposed method introduces a competitive swarm optimizer (CSO) framework tailored for multitasking FS. Specifically, a multitasking construction strategy based on relevance and adaptive threshold is first used to dynamically generate two complementary subtasks: one focusing on a promising feature pool and the other on a global feature pool. The CSO framework enables effective knowledge transfer between these subtasks, improving the overall selection process. Furthermore, an adaptive length variation mechanism is incorporated into the evolutionary process, consisting of two key components: (1) a Gaussian distribution-based variable-length initialization scheme, which enhances the diversity and quality of the initial population; and (2) an adaptive length variation scheme that refines the particle lengths throughout evolution, promoting faster convergence and improved search performance. Extensive experiments conducted on 14 high-dimensional datasets demonstrate that EMT-ALV consistently outperforms several state-of-the-art FS algorithms, achieving better classification accuracy with relatively reduced computation time.