Feature selection is a pre-processing procedure of choosing the optimal feature subsets for constructing model, yet it is difficult to satisfy the requirements of reducing number of features and maintaining classification accuracy. Towards this problem, we propose novel multi-objectives large-scale cooperative coevolutionary algorithm for three-objectives feature selection, termed MLFS-CCDE. Firstly, a cooperative searching framework is designed for efficiently and effectively seeking for the optimal feature subset. Secondly, in the framework, three objectives, feature’s number, classification accuracy and total information gain are established for guiding the evolution of features’ combination. Thirdly, in framework’s decomposition process, cluster-based decomposition strategy is elaborated for reducing the computation; in framework’s coevolution process, dual indicator-based representatives are elaborated for balancing the representative solution’ convergence and diversity. Finally, to verify framework’s practicability, a heart disease diagnosis system based on MLFS-CCDE framework is constructed in cardiology. Numerical experiments demonstrate that the proposed MLFS-CCDE outperforms its competitors in terms of both classification accuracy and metrics of features’ number.

特征选择是选择最优特征子集进行模型构建的预处理过程，但很难满足减少特征数量和保持分类精度的要求。针对这一问题，我们提出了一种新颖的多目标大规模协同协进化算法，用于三目标特征选择，称为MLFS-CCDE。首先，设计了一种协作搜索框架，以高效地寻求最佳特征子集。其次，在框架中，建立了三个目标，即特征数量，分类精度和总信息增益，以指导特征组合的演化。第三，在框架的分解过程中，阐述了基于簇的分解策略，以减少计算量。在框架的协同进化过程中，精心设计了基于双指标的代表，以平衡代表解决方案的收敛性和多样性。最后，为验证该框架的实用性，构建了基于MLFS-CCDE框架的心脏病诊断系统。数值实验表明，本文提出的MLFS-CCDE在分类准确度和特征数量指标方面均优于竞争对手。