Researches On Feature Selection Methods For Functional Data

Since the 21 st Century,the era of big data comes with the sufficient utilization of information resources,the continuous development of computer technology and data mining.In practice,datasets are developing towards large-scale direction and showing an exponential growth trend.This growth includes not only the expansion of data size,but also the diversification of data presentation.Functional data(FD)is one of common data types with rich information.As the high feature dimension of functional data,it is essential to perform feature selection in data analysis,which extracts the relevant information and eliminates the redundant information,so as to achieve the data classification more quickly and accurately.Feature selection of functional data is a process of selecting those features that are slightly correlated and strongly representative from huge function information,which aims to simplify the calculation and improve the generalization ability.Feature selection base on mutual information can express the relationship between data quantificationally,so it is widely used in feature selection.However,the feature selection method based on mutual information existed so far has some shortages,for example,the traditional feature selection methods select features from original dataset directly,the outcomes of feature selection based on different search strategies vary from one to another,which leads to the instability of the results,moreover,it has a higher time complexity of learning algorithms due to all the features traversing during the process of feature selection.If these methods are directly used for functional data feature selection,such problems will be more prominent,and there will be worse accuracy for the classification result of the feature selection.Therefore,this thesis carries out a study on the feature selection method for the functional data according to the existing problems,and the concrete work is detailed as follows.(1)Dynamic mutual information(DMI)feature selection method and dynamic condition mutual information(DCMI)feature selection method are proposed.These two methods overcome the instability of the feature selection.The classification results of classifier are directly applied to the next iteration during the dynamic process of selecting features.By setting the termination conditions in feature selection,the feature subsets which achieve more satisfactory classification results can be obtained with less iteration.(2)Fast feature selection(FFS)method is proposed.In order to simplify the calculation and improve the efficiency of feature selection,this thesis presents a method of feature selection which combines principal component analysis(PCA)and the minimum convex hull method for the feature selection of functional data.This method can obtain a stable subset of features quickly and conquer the problem of time-consuming during feature selection.In addition,the correlations between the features can be fully considered by serving the results of FFS method as an initial subset of features for other methods.Therefore,this thesis combined FFS method with conditional mutual information(CMI)method.By doing contrast experiments,selection strategies are put forward under different time requirements or classification accuracy.Aiming at the feature selection problems of functional data,two feature selection methods based on dynamic mutual information and fast feature selection are presented.These methods can not only ensure the accuracy of classification,but also improve the efficiency of feature selection process and stability of feature selection result.The achievements of the paper enrich the study of the feature selection method for functional data,and have a definite application value.