| Feature selection is an effective high-dimensional data preprocessing method,which aims to remove redundant and irrelevant information in the data,reduce the dimension of the data and the computational complexity of the learning algorithm.How to select accurate features from many features has become a hot topic in artificial intelligence.Most current heuristic feature selection techniques focus on increment or deletion of individual features with respect to a subset of candidate features,which means that information such as collaborative contribution or correlation between features may be lost.Secondly,as the feature dimension continues to increase,the computational complexity of feature selection also increases.Clustering-based feature selection methods can not only eliminate the redundancy of feature subsets to a certain extent,but also reduce some complexity when dealing with high-dimensional data.Therefore,the research on feature selection method based on feature clustering is one of the necessary research directions driven by big data.Most of the existing feature selection driven by feature clustering do not consider the label information when characterizing the correlation between features,which will lead to poor feature clustering effect.To solve this problem,this paper studies the feature relevance measure considering decision information,and performs feature clustering and selection based on this measure.Firstly,a feature selection method based on decision dependent clustering(RDCFS)is proposed.Based on the neighborhood rough set model,an incremental measure of feature joint dependence is designed to describe the redundancy and relevance of data features at the level of classification and identification.On this basis,an evaluation criterion for the optimal feature cluster structure and the optimal graph cut partition of the feature redundant graph are constructed.Finally,a feature centrality and feature dependency measure based on cluster information are given to guide feature selection for high-dimensional data.Considering that the decision-dependent relevance measure of features can only describe the redundancy or relevance of features,based on the idea of ensemble clustering,a feature selection method of ensemble feature fuzzy redundancy clustering(FECFS)was further proposed.According to fuzzy rough set,information theory and other relevant knowledge,the feature correlation of different decision driven is described.At the cluster level,the concept of super feature cluster is proposed,and the calculation method of cluster post progress is given.On the basis of the proposed structure evaluation function,an ensemble clustering method based on super feature cluster fusion is constructed.Finally,a feature importance measure in the cluster is used to guide the feature selection of the data.After the feature selection methods of decision-dependent clustering and ensemble clustering are proposed,in order to verify the effectiveness of the proposed method,the proposed algorithm is compared with the existing feature selection experiments on 8 UCI data sets.The experimental results show that the proposed feature selection method can obtain a more compact feature subset.And the classification performance is better than a variety of existing latest methods.Finally,a single metric is used for feature selection,and the analysis accuracy verifies the necessity of integration. |
