Improved Fisher Score And Hyper-heuristic Differential Evolution Feature Selection Method

With the continuous emergence of data in people's production and life,the volume of data is constantly upgrading,the fields generated are more and more wide,and the dimensions are also higher and higher.This poses unprecedented challenges to the effective methods of analyzing these data,and thus produces the concept of "dimension curse".Feature selection is an effective method to solve dimension problems.Feature selection is becoming more and more important.The role is that after the 1990 s,various feature selection methods began to emerge,and many new methods and ideas emerged endlessly,which played a key role in the development of data dimension reduction technology.Fisher Score,an evolutionary feature selection algorithm based on Fisher discriminant,is a simple and fast feature selection method,which is outstanding in many fields such as face recognition.Firstly,considering that the traditional Fisher Score can not measure the difference between classes completely in formula,the idea of considering the characteristics of different distribution data is added.Secondly,the coefficient is adjusted so that the overlapping relationship between the two classes is considered.Finally,the maximum mutual information coefficient is used to modify the coefficient.The experimental results show that the improved Fisher Score has better recognition ability for features and can measure the more important features for classifiers.However,this filtering method of single variable feature selection has some limitations,that is,it can not measure the redundancy between features,and it also excludes features that may be very important for classification although the score is very low.Therefore,starting from the essence of feature selection-feature subset optimization,this paper introduces a hyper-heuristic search method and uses differential evolution algorithm as a high-level heuristic.Hair-style guidance,design the fitness function to measure the redundancy of the current feature subset for feedback,and select 10 heuristic strategies for low-level design.Because of the retrospective nature of the algorithm,each iteration takes into account the redundancy of features.Finally,a comparative experiment on 14 data sets is carried out.The experimental results show that the hyper-heuristic differential evolution feature selection algorithm is effective.It has good generalization performance.At the same time,because of the high-level heuristic guidance,the low-level heuristic strategy has directionality,which reduces the randomness of heuristic search to a certain extent.