Research On Feature Selection Algorithm Based On Multi-objective Optimizatio

As an effective dimensionality reduction technology,feature selection is one of the important stages of machine learning and data mining.The purpose is to remove irrelevant and redundant data from the original feature data set,so as to further improve data analysis later.In this paper,the feature selection problem will be solved by multiobjective optimization,taking the number of features and classification accuracy as two independent objectives and optimizing them at the same time.Evolutionary algorithm is an effective method to solve multi-objective optimization problems,especially the multi-objective evolutionary algorithm based on decomposition(MOEA/D),which can decompose complex multi-objective optimization problems into multiple scalar subproblems and then solve them,which has better search ability,and it is easier to get an excellent feature subset.Therefore,this paper proposes two feature selection algorithms under the framework of MOEA/D.The main work of this paper is as follows:1.A feature selection algorithm using decomposition based adaptive mutation strategy.The traditional feature selection algorithm uses random mutation probability in the process of population evolution,which cannot select important features well.Therefore,this paper suggests an adaptive mutation strategy for the mutation process in the process of population evolution.During population mutation,this strategy can retain important features with higher probability and delete redundant features with higher probability,so as to obtain feature subsets with high-ranking accuracy and low redundancy.Therefore,a feature selection algorithm using decomposition based adaptive mutation strategy is proposed and applies it to learning to rank,and then compares it with other algorithms on LETOR benchmark data set.The experimental results show that the proposed algorithm can achieve good ranking results with fewer features.2.A feature selection algorithm using decomposition based adaptive neighborhood replacement strategy.The traditional MOEA/D algorithm divides the problems into sub-problem by calculating the Euclidean distance between all weight vectors,and other sub problems in the neighborhood are used to complete coevolution.However,in the process of evolution,the use of fixed-scale neighborhoods will decrease the efficiency of population evolution and affect the classification performance.Therefore,this paper suggests an adaptive neighborhood replacement strategy,which can dynamically provide the neighborhoods of different sizes for the population at different periods of population evolution,so as to balance the diversity and convergence of the population and improve the evolutionary efficiency,to get a better feature subset.Based on this,this paper proposes a feature selection algorithm using decomposition based adaptive neighborhood replacement strategy,then compares it with the traditional evolutionary algorithm on UCI datasets.The experimental results show that the proposed algorithm can obtain fewer features and lower classification error rate.