| Feature selection is an important data pre-processing technique that aims to remove redundant or irrelevant features from the data in order to reduce the dimensionality of the data and improve the performance of the model.In the field of machine learning and data mining,feature selection has been widely used to extract clean and understandable data to build simpler and more efficient models.Therefore,feature selection is often considered as a multi-objective optimization problem i.e.,minimizing the classification error and minimizing the number of selected features.Feature selection is essentially a combinatorial optimization problem,so it is difficult to find the optimal solution in a limited time using general methods,and evolutionary algorithms are now commonly used to quickly search for a set of approximate optimal solutions.Evolutionary algorithms have been shown to be effective in solving multi-objective feature selection problems,but as the dimensionality of the data increases,the search becomes considerably more difficult and many existing evolutionary algorithms suffer from slow convergence and falling into local optimum solutions.Evolutionary algorithms are population-based metaheuristics,and the quality of the initial population is an important factor in the performance of the algorithm.An inappropriate initial population can adversely affect the convergence speed of an evolutionary algorithm,or even reduce the exploration capability of the algorithm to a local optimum.Search operators are also one of the important factors affecting the search ability of the algorithm.Many crossover operators have different search characteristics and search methods,but for feature selection problems on datasets with different characteristics these operators are not robust or even fall into local optima.In order to solve the above problems,we investigate the initial population and search operator of multi-objective evolutionary algorithm,aiming to design new initialization methods and search operators to improve the search ability and convergence speed of multi-objective evolutionary algorithm.The main research of this thesis is as follows:(1)In order to ensure that the initial population is widely distributed in the target space,less similar in the decision space,and to improve the classification performance of the initial feature subset as much as possible,this thesis proposes an initialization method based on interval and mutual information.The method divides intervals according to the number of selected features,and calculates the similarity of solutions in different intervals by Jaccard similarity,and initializes more solutions for intervals with less similarity to ensure that the initial population has better variance and distribution.In addition,a method to determine redundant features is designed based on mutual information in order to reduce redundant features during initialization to improve the classification accuracy of the initial solution.The experimental results show that the initialization method based on partitioning and mutual information can accelerate the convergence of the algorithm on most data sets and enable the algorithm to obtain better performance compared with the traditional random initialization and the initialization methods proposed in the last two years.(2)In order to speed up the evolutionary algorithm when solving large-scale feature selection problems,the adaptive crossover operator is proposed in this thesis.This operator can dynamically determine the number of non-zero genes in the offspring according to the similarity of the parents at different stages of evolution to speed up the convergence of the algorithm,and it combines the feature weights generated by the Relief F method to guide the offspring to select features to improve the quality of the offspring solutions.The experimental results show that the proposed algorithm has better search ability and convergence speed than traditional and state-of-the-art multi-objective evolutionary algorithms,which can converge quickly on most of the data sets and ensure good classification performance while reducing a large number of features. |
