多目标优化在特征选择子集评价中的应用  

Application of multi-objective optimization in feature selection subset evaluation

作　　者：万红[1,2,3] 李蒙蒙 王昊锋[1,2,3] 岳彩通 王力 尚志刚 Wan Hong;Li Mengmeng;Wang Haofeng;Yue Caitong;Wang Li;Shang Zhigang(School of Electrical Engineering,Zhengzhou University,Zhengzhou 450001,China;Industrial Technology Research Institute,Zhengzhou University,Zhengzhou 450001,China;Henan Key Laboratory of Brain Science&Brain-Computer Interface Technology,Zhengzhou 450001,China)

机构地区：[1]郑州大学电气工程学院,郑州450001 [2]郑州大学产业技术研究院,郑州450001 [3]河南省脑科学与脑机接口技术重点实验室,郑州450001

出　　处：《计算机应用研究》2020年第8期2320-2323,2337,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(U1304602,61673353)。

摘　　要：特征选择是处理高维大数据常用的降维手段,但其中牵涉到的多个彼此冲突的特征子集评价目标难以平衡。为综合考虑特征选择中多种子集评价方式间的折中,优化子集性能,提出一种基于子集评价多目标优化的特征选择框架,并重点对多目标粒子群优化(MOPSO)在特征子集评价中的应用进行了研究。该框架分别根据子集的稀疏度、分类能力和信息损失度设计多目标优化函数,继而基于多目标优化算法进行特征权值向量寻优,并通过权值向量Pareto解集膝点选取确定最优向量,最终实现基于权值向量排序的特征选择。设计实验对比了基于多目标粒子群优化算法的特征选择(FS_MOPSO)与四种经典方法的性能,多个数据集上的结果表明,FS_MOPSO在低维空间表现出更高的分类精度,并保证了更少的信息损失。Feature selection is a common dimension reduction approach for processing high-dimensional big data,but it often involves multiple conflicting feature subsets evaluation objectives which are difficult to balance.To reach a compromise among various feature subset evaluations in feature selection and optimize the performance of subset,this paper proposed a subset evaluation multi-objective optimization based feature selection framework and focused on the application of multi-objective particle swarm optimization(MOPSO)in feature subset evaluation.The framework used sparsity,classification ability and information loss to design multi-objective optimization functions.Then it optimized the weight vectors of the features based on multi-objective optimization algorithm,and selected the"knee"of Pareto solution set as optimal vector.Finally,the framework realized feature selection based on weight vector ranking.This paper designed experiments to compare the performance of MOPSO based feature selection(FS_MOPSO)with four classical methods.The results on several standard data sets show that,FS_MOPSO shows higher classification accuracy in low dimensional space while ensuring less information loss.

关 键 词：特征选择 多目标优化 粒子群优化 稀疏 分类 信息损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]