Liu Jianhua, Chen Yuxiang, Li Shanglong
School of Artificial Intelligence, Xiamen Institute of Technology, Xiamen 361021, China.
School of Computer Science and Mathematics, Fujian University of Technology, Fuzhou 350118, China.
Biomimetics (Basel). 2025 May 13;10(5):315. doi: 10.3390/biomimetics10050315.
High-dimensional feature selection is one of the key problems of big data analysis. The binary particle swarm optimization (BPSO) method, when used to achieve feature selection for high-dimensional data problems, can get stuck in local optima, leading to reduced search efficiency and inferior feature selection results. This paper proposes a novel BPSO method with manta ray foraging learning strategies (BPSO-MRFL) to address the challenges of high-dimensional feature selection tasks. The BPSO-MRFL algorithm draws inspiration from the manta ray foraging optimization (MRFO) algorithm and incorporates several distinctive search strategies to enhance its efficiency and effectiveness. These search strategies include chain learning, cyclone learning, and somersault learning. Chain learning allows particles to learn from each other and share information more effectively in order to improve the social learning ability of the population. Cyclone learning introduces a gradual increase over iterations, which helps the BPSO-MRFL algorithm to transition smoothly from exploratory searching to exploitative searching, and it creates a balance between exploration and exploitation. Somersault learning enables particles to adaptively search within a changing search range and allows the algorithm to fine-tune the selected features, which enhances the algorithm's local search ability and improves the quality of the selected subset. The proposed BPSO-MRFL algorithm was evaluated using 10 high-dimensional small-sample gene expression datasets. The results demonstrate that the proposed BPSO-MRFL algorithm achieves enhanced classification accuracy and feature reduction compared to traditional feature selection methods. Additionally, it exhibits competitive performance compared to other advanced feature selection methods. The BPSO-MRFL algorithm presents a promising approach to feature selection in high-dimensional data mining tasks.
高维特征选择是大数据分析的关键问题之一。二元粒子群优化（BPSO）方法在用于解决高维数据问题的特征选择时，可能会陷入局部最优，导致搜索效率降低和特征选择结果不佳。本文提出了一种具有蝠鲼觅食学习策略的新型BPSO方法（BPSO-MRFL），以应对高维特征选择任务的挑战。BPSO-MRFL算法从蝠鲼觅食优化（MRFO）算法中获得灵感，并融入了几种独特的搜索策略，以提高其效率和有效性。这些搜索策略包括链式学习、气旋学习和翻跟斗学习。链式学习使粒子能够相互学习并更有效地共享信息，从而提高群体的社会学习能力。气旋学习引入了随迭代逐渐增加的过程，这有助于BPSO-MRFL算法从探索性搜索平稳过渡到利用性搜索，并在探索和利用之间建立平衡。翻跟斗学习使粒子能够在不断变化的搜索范围内进行自适应搜索，并允许算法对所选特征进行微调，从而增强算法的局部搜索能力并提高所选子集的质量。使用10个高维小样本基因表达数据集对提出的BPSO-MRFL算法进行了评估。结果表明，与传统特征选择方法相比，提出的BPSO-MRFL算法实现了更高的分类准确率和特征约简。此外，与其他先进的特征选择方法相比，它表现出具有竞争力的性能。BPSO-MRFL算法为高维数据挖掘任务中的特征选择提供了一种有前途的方法。
