融合差分进化与分散觅食策略的黏菌算法并应用在高维特征选择上

Slime Mould Algorithm Integrating Differential Evolution and Dispersive Foraging Strategy for High-Dimensional Feature Selection

摘要: 在基因数据这类高维小样本数据的分类任务中，特征选择已成为提升模型泛化能力与可解释性的关键预处理环节。黏菌算法作为一类新型群智能优化方法，具备结构简洁、易于实现的优点。但在处理复杂优化问题时面临探索与开发能力失衡、易陷入局部最优等局限。为此，本文提出一种融合差分进化机制与分散觅食策略的增强型黏菌算法(DBSMA)。该算法通过引入差分进化的变异与交叉操作增强种群多样性，并设计自适应分散觅食策略以扩大全局搜索范围，从而实现算法在全局探索与局部开发之间的有效平衡。在CEC2017标准测试集上的对比实验表明，DBSMA在求解精度、收敛速度与鲁棒性方面均优于12种主流算法与16种改进算法。本文还构建其二进制版本bDBSMA并应用于9个高维医学数据集的特征选择任务，实验结果显示，bDBSMA在显著降低所选特征数量的同时，还能保持较高的分类精度，验证了bDBSMA在高维特征选择任务中的有效性与实用价值。

Abstract: In classification tasks involving high-dimensional, small-sample data such as genomic data, feature selection has become a critical preprocessing step for improving model generalization and interpretability. As a novel swarm intelligence method, the slime mould algorithm is characterized by its simple structure and ease of implementation. However, when addressing complex optimization problems, it faces limitations such as imbalance between exploration and exploitation, as well as a tendency to become trapped in local optima. To address these issues, this paper proposes an enhanced slime mould algorithm, named DBSMA, which integrates a differential evolution mechanism with a dispersed foraging strategy. The algorithm enhances population diversity by incorporating the mutation and crossover operations of differential evolution, and designs an adaptive dispersed foraging strategy to expand the global search range, thereby achieving an effective balance between global exploration and local exploitation. Comparative experiments on the CEC2017 benchmark test suite demonstrate that DBSMA outperforms 12 mainstream algorithms and 16 improved algorithms in terms of solution accuracy, convergence speed, and robustness. Furthermore, a binary version, bDBSMA, is constructed and applied to feature selection tasks on nine high-dimensional medical datasets. Experimental results show that bDBSMA significantly reduces the number of selected features while maintaining high classification accuracy, verifying its effectiveness and practical value in high-dimensional feature selection tasks.