Research On Feature Selection And Classification Method Based On Swarm Intelligence Optimization Algorithm

In most pattern recognition and data mining tasks,Feature Selection(FS)is a key preprocessing step,which helps to avoid the serious impact of irrelevant and redundant features on the performance of classification modes.Feature selection problem is a multi-objective optimization problem,which aims to reduce the number of selected features and ensure the accuracy of classification.In fact,for a given data set with a large number of features,it is often difficult to find a good solution by conventional methods.This thesis studies the feature selection method based on Swarm Intelligence(SI)optimization algorithm,and proposes several improved SI algorithms to improve its ability to solve the feature selection problem.The improved strategy is based on the wrapper feature selection method of linear weighted single objective and multi-objective framework,and combined with the filter feature selection method to guide,in addition,a new transfer function is studied to realize the mapping from continuous search space to discrete search space.(1)An improved Henry Gas Solubility Optimization Based on Stochastic Fractal Search(SFS-HGSO)is proposed.Used for feature selection,function optimization and engineering optimization.Three stochastic fractal strategies based on Gaussian walk,Levy flight and Brownian motion are adopted to conduct diffusion based on high-quality individuals obtained by the original algorithm.Individuals with different fitness are allocated different energies,and the amount of diffusion is determined according to individual energy.This strategy increases the diversity of search strategies,enhances the ability of algorithm exploitation and exploration,and improves the shortcomings of the original HGSO location update method which is single and slow in convergence speed.The proposed algorithm is used to solve the feature selection problem after threshold discretization.It is compared with other swarm intelligent optimization algorithms and proposed feature selection methods on20 standard UCI datasets,and KNN classifier is used to evaluate the effectiveness of the selected features.The algorithm is also used to solve test functions and engineering optimization problems.The experimental results show that these three improvement strategies can effectively improve the performance of HGSO,and obtain excellent results in the problems of feature selection.(2)A Relief F Guided Novel Binary Equilibrium Optimizer(RG-NBEO)is proposed for feature selection.Based on BEO,two new mechanisms are used to improve the algorithm performance.First,based on the concept of opposition-based learning,two new transfer functions,SSr and VVr,are proposed to convert the continuous search space into binary search space,and achieve a good balance between exploration and exploitation.Secondly,a Relief F guidance strategy is proposed to add and delete features in the iterative process according to feature weights.Firstly,the proposed strategy is compared with the classical S-shaped and Vshaped transfer function EO variant,and the best performance BEO variant is selected.The optimal variant is then compared with five excellent swarm intelligence optimization algorithms and six classical filtering feature selection algorithms.The performance of the proposed method was tested on 18 high-dimensional datasets,and the results of different algorithms were statistically evaluated using the Wilcoxon rank sum test and the Freidman rank sum test.The results show that this method can effectively improve the classification accuracy in most cases.(3)A multi-objective optimization algorithm composed of EO and NSGA-Ⅲ is proposed to solve the feature selection problem of high-dimensional data.Three types of transfer functions,S-shaped,V-shaped and U-shaped,are used to convert real coding to binary coding to solve discrete problems,and the influence of three transfer functions on feature selection is compared.In addition,the proposed algorithm also optimizes the binary population structure by constructing external archives,and realizes the selection and optimization of external archive individuals by clustering strategy.The simulation test was conducted on 18 medium and high dimensional data in two groups,and KNN classifier evaluated the classification effect of the model.The first group analyzes the optimization effect of the proposed framework under three transfer functions.The second group of experiments selected the winning algorithm in the first group of experiments and compared it with nine classical multi-objective optimization algorithms such as NSGA-Ⅲ.The evaluation criteria include two optimization objectives of feature selection problem and two optimization indexes of HV and IGD multi-objective problem.The first group of experiments shows that the U3 transfer function performs best in the problem of selective selection;The simulation results compared with other multi-objective optimization algorithms also verify the effectiveness of the proposed strategy.