摘要：

特征选择本质上是一个多模态多目标优化问题, 然而, 多模态多目标优化算法在直接应用于特征选择时, 常常面临冗余特征过多和算法收敛困难等问题。为此, 提出一种决策空间多样性增强的两阶段多模态多目标粒子群优化特征选择(TSMMOPSO-FS)算法。在第一阶段, 提出一种融合互信息的逐层学习领导粒子选择策略。该策略在保持算法收敛性和分类精度的同时, 能够有效地剔除冗余特征。在第二阶段, 引入一种基于距离多样性和特征频率的粒子速度奖惩机制以提高决策空间的多样性。该机制通过对全体粒子速度进行自适应调整, 使粒子能够更广泛地探索不同的解空间, 提高搜索等价特征子集的能力, 避免陷入局部最优。在8个不同规模的UCI数据集上和5种多模态多目标优化算法进行对比实验, 实验结果表明, 该算法不仅表现出更优秀的收敛性和多样性, 在不损失分类正确率的同时也搜索到了更多的等价特征子集。

关键词: 特征选择, 多模态多目标优化, 粒子群优化算法, 互信息, 决策空间多样性增强

Abstract:

Feature selection is inherently a multimodal, multiobjective optimization problem. However, multimodal multiobjective optimization algorithms directly applied to feature selection often face challenges such as excessively redundant features and difficulties in algorithm convergence. To address these issues, a Two-Stage Multimodal MultiObjective Particle Swarm Optimization Feature Selection (TSMMOPSO-FS) algorithm that enhances decision space diversity is proposed. In the first stage, the algorithm introduces a layer-by-layer learning leader particle selection strategy that integrates mutual information. This strategy effectively eliminates redundant features while maintaining algorithm convergence and classification accuracy. In the second stage, a particle velocity reward-punishment mechanism based on distance diversity and feature frequency is introduced to enhance the diversity of the decision space. This mechanism adaptively adjusts the velocity of all particles, enabling a broad exploration of different solution spaces, improving the ability to search for equivalent feature subsets, and avoiding local optima. Comparative experiments conducted on eight UCI datasets of varying sizes using five multimodal multiobjective optimization algorithms reveal that the proposed algorithm not only exhibits superior convergence and diversity but also identifies more equivalent feature subsets without compromising classification accuracy.

Key words: feature selection, multi-modal multi-objective optimization, Particle Swarm Optimization (PSO) algorithm, mutual information, decision space diversity enhancement