Research On Feature Selection Method Based On Differential Evolution Algorithm

With the development of network technology,data has become an available capital,which has brought opportunities and challenges to the development of data mining technology.In the real world,data is inherently complex.And the large number of features adds complexity to the challenge.So,how to remove redundant data and retain related data has become a research hotspot for data mining and machine learning tasks.In practical tasks,in order to avoid the data dimension disaster problem and remove irrelevant features to reduce the difficulty of data processing tasks,feature selection technology is used as a preprocessing process that can retain semantics.Feature selection is mainly based on the distribution characteristics of the data,using an appropriate search mechanism and fitness function to select an optimal or similar feature subset from the original feature set,thereby improving time,space and data analysis resultsThe research of feature selection algorithm is mainly divided into two aspects: search strategies and evaluation criteria.In this paper,information theory and Pearson correlation coefficient correlation technology are used as evaluation criteria,and improved differential evolution algorithm optimization method is used as a search mechanism to optimize feature selection.The main work is as follows:(1)Proposed an improved adaptive differential evolution for feature selection algorithm(ISHADEFS).First,in terms of feature selection evaluation criteria,in order to improve the influence of the evaluation criteria and the accuracy of feature selection,a fitness function was designed and implemented to enhance the effect of mutual information and Pearson correlation coefficient.Secondly,in order to solve the contradiction between population diversity and convergence in the differential evolution algorithm,a triangular mutation operator is proposed.A threshold-based mutation strategy is designed,and two mutation operators are selected through the threshold.Finally,experimental verification is performed.By analyzing the algorithm fitness function curve,the best threshold of the mutation strategy and the influence factor of the fitness function are obtained.The ISHADEFS algorithm is compared with existing feature selection methods on different classifiers.The experimental results show that the ISHADEFS algorithm proposed in this paper can evolve fewer features and obtain better classification results.(2)Proposed an improved multi-objective differential evolution for feature selection algorithm(IMODEFS).First,feature selection is modeled as a multi-objective optimization problem,and consider the correlation and similarity between features to deal with unwanted features.Secondly,in order to improve the diversity of the population,a cooperative mutation strategy was designed.By setting thresholds,the “current to pbest” mutation operator with fast convergence speed can perform a cooperative mutation operation with DE/rand/1 to better select the mutation operator.Finally,the Pareto front of the IMODEFS algorithm is proved by experiments to prove that the algorithm can provide sufficient population Pareto fronts.The experimental comparison between IMODEFS and existing feature selection algorithms proves that the IMODEFS algorithm proposed in this paper is superior in classification performance and optimization performance.