Big data has been mentioned more and more in recent years.The development of science,technology and the rapid flow of information have made the rapid development of contemporary society an era of data explosion.Feature selection has been widely studied in classification for improving classification accuracy and reducing computational complexity,and it has been widely used in the field of data mining.In recent years,evolutionary computation(EC)methods have become a kind of important methods to solve feature selection problems.The random search strategies of the EC methods have a strong ability to search for solutions.However,as the sizes of datasets increase,more and more irrelevant and redundant features appear.These irrelevant and redundant features may lead to local optimal problems on large-scale feature selection problems.At the same time,when using the EC method to solve large-scale feature selection problems,only one candidate solution generation strategy(CSGS)and fixed parameters of traditional EC methods do not perform well on the search optimal subset.In order to better solve the problem of large-scale feature selection,in this paper we mainly do the following research work:(1)To solve the problem of insufficient adaptability caused by single strategy and fixed parameters in existing EC methods,we propose a particle swarm optimization algorithm(SPS-PSO)based on self-adaptive parameter and strategy and use it to solve large-scale feature selection problems.In SPS-PSO,the strategy self-adaptive mechanism and the parameter self-adaptive mechanism are employed to the framework of particle swarm optimization(PSO).In addition,Furthermore,in order to examine the effects of the final subset of features that are generated when different classifiers are used as evaluation functions for feature selection,k-nearest neighbor(KNN),latent dirichlet allocation(LDA),extreme learning machine(ELM)and support vector machine(SVM)are used as evaluation functions for the feature selection.(2)To solve the difficult problem of large-scale feedforward neural network(FNN)optimization problem,based on the proposed SPS-PSO algorithm,we applies SPS-PSO and the SPS-PSO based feature selection method to large-scale FNN optimization problems,designed to reduce the complexity of large-scale FNN optimization problems while improving accuracy.First,we use SPS-PSO to optimize FNN problems directly on the original dataset.Then,the original dataset is generated by the SPS-PSO-based feature selection method to generate a smaller feature subset,and then the feature subset is used as the input of FNN problem,.(3)To solve the problem that the CSGSs of SPS-PSO algorithm has a single source of strategy,we introduce three new CSGSs from the differential evolution(DE)algorithm while reserving three CSGSs from PSO to form a new strategy pool with six CSGSs and form a new parameter and strategy self-adaptive differential particle swarm optimization algorithm(SPSDPS).We apply this algorithm to the intrusion detection problem of wireless sensor networks(WSN)and verify its performance through datasets. |