Research And Improvement Of Feature Selection Algorithms Based On Evolutionary Learning

Feature selection is one of the earliest branches in machine learning field.It is still a very important process of data preprocessing and one of the commonly used data dimensionality reduction methods.It is widely applied to a series of practical problems,such as classification,dimension reduction,handwriting recognition,etc.Feature selection refers to the feature subset that makes some evaluation criteria optimal from the selection of the original feature set,whose final goal is to remove irrelevant and redundant features according to some selection criteria,so as to select the smallest feature subset that makes tasks such as classification,regression and so on achieve similar or even better results,and improves generalization ability of the algorithm.Evolutionary algorithms are bionic algorithms that simulate the natural genetic evolution rule.It is especially suitable for dealing with highly complex nonlinear problems which are difficult to be solved by traditional search methods.As evolutionary algorithms are reported to be suitable for optimization problems,the application of evolutionary algorithms(EA),particle swarm optimization(PSO)and ant colony algorithm(ACO)in the field of feature selection has achieved satisfactory results.In order to better solve the problem of feature selection,Manizheh Ghaemi et al.proposed a new,efficient FSFOA algorithm which is good at global searching ability.Compared with other algorithms,the FSFOA algorithm usually achieves higher classification accuracy and has better generalization performance with only a smaller computational cost.In spite of this,there are still some shortcomings.In the FSFOA method,the randomness of initialization phase,the limitations of updating mechanism and the inferior quality of the new tree in the local seeding stage severely limit the classification performance and dimensionality reduction ability of the algorithm.Inspired by the IniPG and MFOA algorithms,this paper improves the shortcomings of the FSFOA algorithm.In the initialization stage of FSFOA algorithm,we adopt the new initialization strategy,use the advantages of forward selection and backward selection and discard their shortcomings to form a two-way selection strategy;On the update mechanism,the limitations of the traditional updating mechanism are overcome,and the dimension reduction is also taken into consideration;In order to avoid too many inferior trees in the forest which increase the search difficulty and affect the classification performance,we adopt the extreme greedy strategy to form a new feature selection algorithm IFSFOA,which minimizes the number of features while maximizing the classification performance.In the improved process,in order to avoid local optima problems caused by greed strategy,The object that global seeding affects is changed from candidate population to candidate population and all the Age-0 trees in the forest but the number of trees used before and after changing in the global seeding stage has not changed in the IFSFOA algorithm,making the same 0-Age trees affect both local seeding and global seeding,which solves the problem of easy to fall into the local optimal solution due to the extreme greedy strategy to a certain extent.In the experimental stage,IFSFOA uses SVM,J48 and KNN classifiers to guide the learning process while utilizing the machine learning database UCI for testing.The results of the experiments show that compared with FSFOA,IFSFOA has a significant improvement in classification performance and dimensionality reduction.Comparing IFSFOA algorithm with more efficient feature selection methods proposed in recent years,IFSFOA is still very competitive in both accuracy and dimensionality reduction.