.png){kind=icon}
.png){kind=icon}
Return
Enhancing high-dimensional feature selection with dynamic tri-competitive swarm optimization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.jocs.2026.102882.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
As a key machine learning preprocessing technique, feature selection (FS) aims to extract a subset of relevant, non-redundant features from high-dimensional data while eliminating irrelevant ones. It reduces dimensionality to alleviate the “curse of dimensionality”, optimizes model efficiency, lowers computational costs, and enhances generalization performance, making it indispensable for high-dimensional tasks in pattern recognition, data mining, and predictive analytics. However, high-dimensional FS faces prominent challenges, exponential feature space expansion causes intractable search space combinatorial explosion, strong feature correlations lead to redundant selection, and traditional optimization algorithms often suffer from premature convergence to local optima, failing to find the global optimal subset. As FS is inherently a combinatorial optimization problem, swarm intelligence algorithms are widely used for their global search advantages. The traditional Competitive Swarm Optimizer (CSO) enhances population diversity via pairwise competition, yet exhibits critical limitations in high-dimensional FS, unbalanced exploration-exploitation trade-off and reliance on a single fitness metric, which induces single-objective bias and fails to meet FS’s dual requirements of high classification accuracy and compact feature subsets. To address these issues, this paper proposes a Dynamic Tri-Competitive Swarm Optimizer (DTCSO), with core contributions as follows: (1) A dual-index-based tri-competition mechanism mitigates single-objective bias by integrating classification accuracy and feature subset size, enhancing selection pressure for high-quality subsets balancing effectiveness and compactness. (2) A dynamic multi-factor adjustment mechanism adaptively tunes control parameters per evolutionary stage, balancing exploration and exploitation for efficient optimization. (3) An integrated spiral search strategy strengthens local search capability, breaking away from local optima to improve subset quality. These components synergistically boost DTCSO’s high-dimensional FS performance. Experiments on 19 datasets from the ASU feature selection repository demonstrate that DTCSO outperforms the comparative algorithms in both classification accuracy and subset compactness. For reproducibility, the source code is publicly available at https://github.com/lsb00/DTCSO .
Keywords:
feature selection
swarm intelligence
combinatorial optimization
high-dimensional data
dynamic tri-competition
Journal
J
IF:
3.7
Papers:
205
Citations:
0