Research On Feature Selection Algorithm Based On Evolutionary Computation

Feature selection can reduce the dimension of data by removing redundant and useless information according to an evaluation criteria,so as to improve the efficiency and performance of the learning algorithm.Therefore,it has important research significance and practical value in the field of information mining and machine learning.The feature selection algorithm based on evolutionary computation(EC)is one of the current research directions in the field of feature selection.The existing EC based feature selection algorithms have achieved a lot of impressive research achievements,but also exist some limitations.Firstly,these researches mostly are the main form of traditional EC algorithm or its variants.However,such algorithms often have problems of low time efficiency and premature convergence,which will affect the optimization process.Secondly,the existing researches mostly focus on improving evolutionary operator,but an excellent objective function can also affect the performance of learning algorithm and the landing application.In view of the above problems,the main work of this paper is as follows:(1)For the convergence problem,an I-Ching with differential evolution based algorithm for feature selection(IDEFS)is proposed in this paper.Firstly,according to the main idea of IChing divination evolutionary algorithm(IDEA)in solving optimization problems,an I-Ching based algorithm for feature selection(IFS)is suggested to find the excellent feature subsets from original feature space.Then the proposed IDEFS algorithm is an enhancement of IFS by introducing differential evolution strategy to accelerate the convergence of algorithm and further improve the searching performance.Finally,experiments checked out with standard dataset from UCI repository reveal that the proposed IDEFS performs better in terms of classification accuracy,precision,recall and feature reduction than the competing feature selection algorithm.Thereby the IDEFS is an effective and promoting algorithm in case of feature selection.(2)For the objective function problem,an emotion-related feature selection algorithm based on EC and normalized mutual information(IDEFS-NMI)is proposed,which introduces normalized mutual information(NMI)into IDEFS to research the feature selection problem on EEG emotion recognition.NMI can filter features that are highly correlated with the class and have low redundancy between features according to the data distribution,so it can be used as an important indicator to measure features in the objective function.Through searching the candidate feature subset of EEG after feature extraction and establishing the emotion classification model,the designed NMI based objective function can evaluate the performance of the candidate solution for selecting the outstanding features related to emotion.The comparative experiments on DEAP data using 5-fold cross validation illustrate that the proposed IDEFS-NMI achieves a higher accuracy of 74.92% for valence and 74.89% for arousal than that of the other competing methods,verifying the practicability and effectiveness of the proposed algorithm.In summary,the proposed feature selection algorithm IDEFS in this paper has good global convergence and can be effectively used in feature selection problems;the further proposed IDEFS-NMI algorithm using the objective function based on normalized mutual information can achieve better recognition results in EEG emotion recognition research with a certain practical application value.