Feature Selection Based On Particle Swarm Optimization Method And Implementation

Feature selection method is widely used in today's "digital" era,and it appears frequently as a means of data preprocessing.The purpose of this technology is not to damage the value information expression ability of data,but also to sharply reduce the amount of data.Accuracy must be the eternal theme of feature selection,and time is also an important evaluation index.And with the increasing number of columns of data to be solved,the expansion of problem space brings great challenges to the algorithm.In this paper,particle swarm optimization(PSO)evolutionary algorithm is used to optimize the feature selection(1)Firstly,this paper analyzes the complementary characteristics of the two patterns in feature selection problem,and proposes a two-stage global optimal value(Gbest)updating method by combining filter and wrapper in a specific way.In the first stage,the simplified symmetric uncertainty function of the filter class is used as the precision function of the evaluation subset,and the individual optimal solution of each particle and the global optimal solution of the community are updated.In the next stage,the most promising particles are obtained by two ways,and the wrapper class evaluation function is used to select the best update global optimal solution.This mechanism can achieve a balance between time and classification effect.Compared with the classical algorithm,the value of the algorithm is proved.(2)In the double Gbest feature selection algorithm based on particle swarm optimization(DGPSO),the update mechanism of particle swarm optimization is changed.By imitating the update strategy of quantum particle swarm optimization,the particle rate formula is eliminated,and the particle position is updated by XOR and mutation probability setting.DGPSO has better global search ability than BPSO.(3)In order to further optimize DGPSO algorithm,its performance in time and feature length is improved.In this paper,based on DGPSO,we propose a new feature selection algorithm,high dimensional double Gbest feature selection algorithm(LDGPSO),which is more suitable for feature selection problems with large dimensions.Before limiting the length,attention feature selection mechanism(AFS)should be used to re rank all attribute columns from good to bad.All features are partitioned according to the ranking,so that the thresholds of attribute features in different regions are different during initialization or update.In other words,the top ones are more likely to be chosen.LDGPSO limits the number of attributes selected by particles according to the execution process and the intermediate results of the algorithm.In order to reduce the length of selection results,reduce the amount of calculation and eliminate redundant attributes.