Research On Multi-objective Feature Selection Based On Improved NSGA-? Algorithm

With the rapid development of the information age and the increasing amount of data,the current era has transformed into the era of big data.Data analysis and processing have become the focus of research.Feature selection is an important tool to improve classification performance,while reducing the complexity of the classifier and reducing the processing time of classification,especially in solving classification problems on high-dimensional data sets.Most of the existing optimization methods take the linear combination of two different indicators as the goal,and transform the feature selection into a single-objective optimization problem.However,for linear combination,a suitable weight is very critical and the searching process is complicated and difficult.In recent years,many researchers have transformed feature selection into multi-objective optimization problem to optimize two objectives at the same time which are classification accuracy and solution size,but these are far from enough.Feature selection model that only optimize these two objectives can solve most of the multi-objective feature selection problems,but it is not suitable for missing data sets.In real life,missing data is common due to various reasons.When applying feature selection on the missing data set,the more missing data is included in the selected features,the less reliable the selected features will be.In order to solve the problem of multi-objective feature selection on missing data sets,this paper includes the following research:(1)Effective crossover operators can prevent premature convergence of evolutionary algorithms and further improve the diversity of offspring populations,but there are few improvements on crossover operators in the existing studies.Therefore,to solve this problem,aNSGA-? with self-adaptive Crossover Operator(SaNSGA-?)is proposed.SaNSGA-? adds an adaptive mechanism,which includes a variety of crossover methods and increases the diversity of offspring populations.This paper collects four excellent crossover operators and puts them into the pool of crossover operators,using Roulette to randomly select crossover operators and updating the selection probability in time.The experimental results show that,compared with the original NSGA-? algorithm and other comparison algorithms,the SaNSGA-? algorithm has better performance according to IGD and HV results,as it has a more uniform distribution of offspring and increases diversity of offspring.(2)It is impossible to select reliable features when making feature selection on missing data sets.For this problem,this paper uses mean interpolation method to interpolate missing data sets into complete data sets,and regards feature selection as a multi-objective optimization problem,so that single-objective feature selection problem is transformed into multi-objective feature selection problem.(3)For multi-objective feature selection problem on missing data sets,in order to choose the more reliable offspring,this paper adds reliability as the third objective to the feature selection model that has two objectives of classification accuracy and solution size,thereby expanding feature selection problem into a three-objective feature selection model and allowing all multi-objective optimization algorithm to perform feature selection on the missing data sets using this model.In addition,the experimental results show that,if the proposed SaNSGA-? algorithm is applied to the model,SaNSGA-? is more suitable for this model,and it can select a solution set with high classification accuracy,fewer features and reliability.