基于影响力社区检测与蚁群算法的特征选择  

Feature selection based on influence community detection and ant colony optimization algorithm

作　　者：叶小艳[1] 叶小莺 周化[1] YE Xiao-yan;YE Xiao-ying;ZHOU Hua(Department of Network Technology,South China Institute of Software Engineering Guangzhou University,Guangzhou 510990,China;Department of Computer Science and Technology,Neusoft Institute of Guangdong,Foshan 528225,China)

机构地区：[1]广州大学华软软件学院网络技术系,广东广州510990 [2]广东东软学院计算机科学与技术系,广东佛山528225

出　　处：《计算机工程与设计》2019年第9期2684-2691,共8页Computer Engineering and Design

基　　金：2017年外经外贸发展专项基金项目(促进服务贸易创新发展项目)(2160699-87);2017年广东省本科高校教学质量与教学改革工程建设基金项目(粤教高涵[2017]214号);广州大学华软软件学院院级质量工程重点建设专业基金项目(ZDZY201701)

摘　　要：针对多变量特征选择算法计算效率低、冗余度高的问题,提出一种基于影响力社区检测与蚁群算法的特征选择算法。计算每对数据点之间的相似性,组成无向图,通过人工蚁群优化算法将网络划分为簇;使用社区检测算法对特征进行分类,选择冗余度最小的特征子集;蚁群初始化阶段通过度量特征与类的相关性,初始化信息素。基于人工合成数据集与标准的公开数据集进行实验,实验结果表明,该算法实现了较高的分类准确率、敏感性、特异性,其计算效率处于可接受范围内。Aiming at the problems of low computational efficiency and high redundancy of multivariate feature selection algorithm,a feature selection algorithm based on influence community detection and ant colony optimization algorithm was proposed.The similarity of each pair of data points was computed to construct a graph,the network was divided into clusters using ant colony optimization algorithm.The community detection was adopted to classify the features,and the feature subset with minimal redundancy was selected.In the initial phase of ant colony,the pheromone was initialized by measuring the correlations between features and clusters.Experimental results based on the synthetic dataset and standard public datasets show that,the proposed algorithm performs good classification accuracy,sensitivity and specificity,at the same time,it realizes reasonable computational efficiency.

关 键 词：社区检测 特征选择 人工蚁群优化算法 多元判别分析 人工智能 数据分析 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]