基于归一化模糊联合互信息最大的特征选择

发布时间：2018-01-13 08:21

本文关键词：基于归一化模糊联合互信息最大的特征选择　出处：《计算机工程与应用》2017年22期 　论文类型：期刊论文

【摘要】：特征选择就是从特征集合中选择出与分类类别相关性强而特征之间冗余性最小的特征子集,这样一方面可以提高分类器的计算效率,另一方面可以提高分类器的泛化能力,进而提高分类精度。基于互信息的特征相关性和冗余性的评价准则,在实际应用中存在以下的问题:(1)变量的概率计算困难,进而影响特征的信息熵计算困难;(2)互信息倾向于选择值较多的特征;(3)基于累积加和的候选特征与特征子集之间冗余性度量准则在特征维数较高的情况下容易失效。为了解决上述问题,提出了基于归一化模糊互信息最大的特征评价准则,基于模糊等价关系计算变量的信息熵、条件熵、联合熵;利用联合互信息最大替换累积加和的度量方法;基于归一化联合互信息对特征重要性进行评价;基于该准则建立了基于前向贪婪搜索的特征选择算法。在UCI机器学习标准数据集上的多组实验,证明算法能够有效地选择出对分类类别有效的特征子集,能够明显提高分类精度。
[Abstract]:Feature selection is to select the feature subset which has strong correlation with the classification category and the least redundancy between the features from the feature set. On the one hand, it can improve the computing efficiency of the classifier. On the other hand, it can improve the generalization ability of classifier, and then improve the classification accuracy. In the practical application, there are the following problems: 1) the probability of the variable is difficult to calculate, and then the information entropy which affects the feature is difficult to calculate. (2) Mutual information tends to select features with more values; In order to solve the above problem, redundancy metrics between candidate features and feature subsets based on cumulative sum are easy to fail in the case of high feature dimension. Based on the normalized fuzzy mutual information maximum feature evaluation criterion, the information entropy, conditional entropy and joint entropy of variables are calculated based on fuzzy equivalence relation. A measure of the cumulative sum of the maximum substitution of joint mutual information; The importance of feature is evaluated based on normalized joint mutual information. Based on this criterion, a feature selection algorithm based on forward greedy search is established. Experiments on UCI machine learning standard data set show that the algorithm can effectively select feature subsets that are effective for classification categories. The classification accuracy can be improved obviously.
【作者单位】：武汉科技大学城市学院实验实训中心;华中科技大学软件学院;
【基金】：国家自然科学基金(No.61472289) 中央高校基本科研业务费专项资金(No.2015IVA067) 河南省教育厅自然科学基金(No.15A110011)
【分类号】：TP18
【正文快照】： 1.武汉科技大学城市学院实验实训中心,武汉4300832.华中科技大学软件学院,武汉4300001.Research and Training Center of City College,Wuhan University of Science and Technology,Wuhan 430083,China2.School of Software Engineering,Huazhong University of ScienceTech

【相似文献】