基于分解的自适应邻域替换策略的特征选择算法  

Feature Selection Algorithm Using Decomposition Based Adaptive Neighborhood Replacement Strategy

作　　者：马璐 苏华[1] MA Lu;SU Hua(School of Computer Science and Technology,Tiangong University,Tianjin 300387,China)

机构地区：[1]天津工业大学计算机科学与技术学院,天津300387

出　　处：《长春师范大学学报》2022年第6期44-48,59,共6页Journal of Changchun Normal University

基　　金：国家自然科学基金项目“面向认知物联网的频谱大数据处理关键技术研究”(61972456);天津市自然科学基金一般项目“面向认知物联网的频谱大数据处理关键技术研究”(20JCYBJC00140);泛网无线通信教育部重点实验室(BUPT)开放课题“面向认知物联网的频谱大数据处理关键技术研究”(KFKT-2020101)。

摘　　要：特征选择是数据处理和机器学习中一个重要的过程,提高分类性能和减少特征数目是特征选择的两个主要目的,因此特征选择可以被定义为一个多目标优化问题。为了进一步在提高特征选择分类性能的同时降低特征数目,本文在MOEA/D的框架下,提出了一种基于自适应邻域替换策略(AR)的特征选择算法,能有效避免使用固定规模的邻域而造成种群进化缓慢从而影响分类性能的问题。将所提出的算法与其他算法在UCI中的5个不同分类数据集上相比,结果表明,本文算法得到比其他方法更优的结果。Feature selection is an important process in data processing and machine learning.Improving classification performance and reducing the number of features are the two main purposes of feature selection.Therefore,feature selection can be defined as a multi-objective optimization problem.In order to further improve the classification performance of feature selection and reduce the number of features,a feature selection algorithm based on adaptive neighborhood replacement(AR)strategy is proposed,which can effectively avoid the problem of slow population evolution caused by the use of fixed-scale neighborhood,which affects the classification performance.Finally,the proposed algorithm is compared with other algorithms on five different classification data sets.The results show that the proposed algorithm obtains better feature subsets than other algorithms.

关 键 词：多目标优化 特征选择 MOEA/D 邻域替换 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]