基于特征聚类的封装特征选择算法

摘要

针对多维数据集,为得到一个最优特征子集,提出一种基于特征聚类的封装式特征选择算法.在初始阶段,利用三支决策理论动态地将原始特征集划分为若干特征子空间,通过特征聚类算法对每个特征子空间内的特征进行聚类;从每个特征类簇里挑选代表特征,利用邻域互信息对剩余特征进行降序排序并依次迭代选择,使用封装器评估该特征是否应该被选择,可得到一个具有最低分类错误率的最优特征子集.在UCI数据集上的实验结果表明,相较于其它特征选择算法,该算法能有效地提高各数据集在libSVM、J48、Naive Bayes以及KNN分类器上的分类准确率.%To obtain an optimal feature subset of multi-dimensional data,a feature selection algorithm based on feature clustering and wrapper (FC_ W) was proposed.In the initial stage,the original feature set was divided into a number of feature subspaces using the three-way decision theory,and the features of each feature subspace were clustered using the feature clustering algorithm.The representative features were selected from each feature cluster,and the remaining features were sorted in descending order and iteratively selected using the neighborhood mutual information (NMI) between them.In this selection process,a wrapper was utilized to evaluate whether the selected feature should be selected or not.An optimal feature subset with a minimum classification error rate was obtained.Experimental evaluation on UCI data sets shows that,compared with the feature selection algorithms in other literatures,the proposed algorithm has higher classification accuracy in libSVM,J48,Naive Bayes and KNN classifiers.