Research On Evolutionary Optimization Algorithms For Solving Feature Selection Problems

With the rapid development of economic level and information technology level,the society has completely entered the era of data.However,these data have high dimensions,and many features of them are useless.These characteristics will cause problems such as low utilization of data storage space,reduced classification performance and low operating efficiency.Therefore,finding the optimal data quickly and efficiently is a problem that must be considered and solved.Feature selection is an important method for data preprocessing.When processing a large amount of data to solve classification problems,it can effectively improve the classification accuracy and reduce the time complexity of the classification problem,thereby improving the performance of machine learning algorithms.In recent years,evolutionary algorithms have been widely used to solve feature selection problems because of their global search capabilities.However,when the dimensionality of the dataset is increased to a certain degree,many irrelevant or redundant features make many methods easily fall into the problems of "the curse of dimensionality" and local optimization.Many existing evolutionary algorithms cannot effectively solve such problems.At present,researchers mostly improve the performance of evolutionary algorithms from the perspective of optimization of parameters and structure of evolutionary algorithms.In order to improve the performance of evolutionary algorithms from different angles and better solve the feature selection problem,especially the large-scale feature selection problem,this paper has mainly done the following research work:(1)Effective population initialization methods can improve the convergence performance of evolutionary algorithms to solve feature selection problems.However,in the existing research,there are relatively few studies on population initialization.Filters are simple and efficient,and can quickly evaluate and select useful features.These methods have not been widely used to improve the performance of population initialization.Therefore,an improved particle swarm optimization algorithm based on filter and threshold selection-based population initialization(FTSI-PSO)is proposed to solve the problem of feature Selection.(2)In order to solve the problems of low universality of FTSI-PSO algorithm and failure to make full use of the interaction between features,the particle swarm optimization with filter and probability-based population initialization is proposed.Experimental results show that the algorithm has better convergence speed and higher classification performance in solving feature selection problems.(3)In order to solve the local optimal stagnation problem when solving large-scale multi-objective feature selection problem,a multi-objective self-adaptive particle swarm optimization algorithm is proposed.The classification accuracy and solution size are taken as two goals.Then,the self-adaptive mechanism,the fast non-dominated sorting method,the crowding distance calculation method and the elite strategy are adopted to build a multi-objective feature selection algorithm based on the PSO algorithm.Experimental results show that the performance of this algorithm is better than other multi-objective algorithms when solving feature selection problems.