Research On Community Detection Based Particle Swarm Optimization Feature Selection Algorithm

With the development of society and the progress of science and technology,the amount of data have already presented the explosive growth,the performance of learning algorithm will decrease with the increase of data set characteristics in the field of machine learning.In recent years,as an effective method to solve this problem,feature selection has been widely concerned,many scholars have developed and improved many feature selection methods,and achieved considerable results.Among these effective algorithms,PSO has become the focus of current researches with the advantages of few parameters,simple operation and strong local search ability.This thesis will further analyze the particle swarm optimization algorithm and feature selection problems,and utilize the hidden information between features obtained by the community detection algorithm to guide population evolution.The main work includes the following two aspects:(1)Due to the particle swarm optimization algorithm is easy to fall into local optimum and it is difficult to find of high-quality feature subset in solving the problem of feature selection.A grouping based particle swarm optimisation algorithm for Feature Selection CBPSOFS is proposed.The general idea of the algorithm is as follows: Firstly,the features is transformed into a feature map,and then the feature is divided into different features communities according to the algorithm of community detection.According to the similar information contained in the same features community,an adaptive updating strategy based on the community is designed to keep the diversity of the population,while balancing the local search and the global search,and effectively overcome the phenomenon of falling into the local optimum prematurely.In addition,an information-gain based initialization strategy and the reset strategy based on historical information are designed to further improve the performance of the algorithm.By comparing the real datasets with the four existing feature selection methods,the effectiveness of the proposed CBPSOFS algorithm is verified.(2)Due to the high-dimensional data set contains a large number of redundant,irrelevant features,which leads to the low performance of feature selection algorithm,a feature selectionalgorithm KBPSOFS for high-dimensional data is proposed.This method proposes the concept of core particles combined with community detection algorithm.When the global optimal value of the population Gbest is unchanged several times in a row,core particles are used instead of Gbest to guide the evolution of the population and enhance the search ability of the algorithm.In addition,when the entire population Pbest has not changed,a population transformation strategy is proposed using the current population non-dominated solution.The purpose is to reset the population and remove irrelevant and redundant features.Through the experiments on real datasets,the results shows that KBPSOFS can effectively reduce the number of features and improve the classification accuracy.