Adaptive mixed mutation for unsupervised feature selection using differential evolution

Xuechen Mu

doi:10.1016/j.neucom.2026.133807

Adaptive mixed mutation for unsupervised feature selection using differential evolution

2026-05-06

0

PRE

AI

DOI：10.1016/j.neucom.2026.133807

Original

Original request for help

Save

Abstract

En 中文

As the availability of big data grows, identifying relevant information within large, unlabeled datasets has become increasingly important. Unsupervised feature selection methods address this challenge without relying on labeled data, but the absence of labels often hinders their ability to uncover informative feature subsets. To overcome this limitation, we propose AMUSE, a novel Differential Evolution (DE)-based algorithm that integrates an adaptive mixed mutation strategy and an assistance module, combined with a multi-indicator fitness function, to enhance solution diversity and uncover more promising feature subsets. First, we enhance three existing mutation strategies by incorporating a “local best” element derived from the crossover space, replacing random components and thereby improving the continuity of high-quality solutions. These enhanced strategies, together with five existing mutation strategies containing “global best” elements, form a candidate pool from which we adaptively select and weight mutation strategies. By applying cosine similarity and a variable decay factor, we create a mixed mutation strategy that maintains low correlation and high diversity. Next, we introduce an assistance module that evaluates whether the newly generated mixed mutation strategy outperforms its predecessor, dynamically adjusting the crossover probability ( ) to retain more promising solutions. This data-driven adjustment preserves high diversity in the mutation space, increasing the likelihood of identifying superior feature subsets. In addition, we design a fitness function integrating multiple indicators to effectively apply AMUSE in unsupervised feature selection scenarios. We conduct extensive experiments on 50 UCI datasets, comparing AMUSE against 22 existing algorithms spanning sparse-based, similarity-based, and evolution-based methods. Our results demonstrate that AMUSE consistently produces more effective feature subsets, especially in binary classification tasks. A deeper analysis of the diversity of generated feature subsets further supports our claims. The source code is publicly available at https://github.com/Mxc666/AMUSE.git .

Keywords:

Unsupervised feature selection

Differential Evolution

Adaptive mixed mutation

Feature subset selection

Fitness function

Adaptive mixed mutation for unsupervised feature selection using differential evolution

Abstract

Journal

Organization

Cited Papers