Research And Improvement Of Feature Selection Algorithm Based On Crow Search Algorithm

Feature selection is a process of selecting the optimal feature subset from the original data set.It is an important means to improve the performance of learning algorithm by reducing the dimension of the data set.It is also a key data preprocessing step in the process of machine learning.On the premise of ensuring a certain classification accuracy,the feature selection method is used to delete the irrelevant and redundant features in the data set,so as to solve the problem of large number of features in the data set and complex interaction between features,thus reducing the difficulty in subsequent machine learning tasks and making the classifier work fast and accurately.Meta-heuristic algorithm is a kind of general heuristic algorithm.Under the condition of limited time and cost,the meta-heuristic algorithm is helpful to find the approximate optimal solution from the huge solution space in the limited time.Scholars try to apply meta-heuristic algorithm to solve complex feature selection problems.Among them,the meta-heuristic algorithm inspired by nature has proved its potential in various fields and can provide solutions for different optimization problems.However,no specific algorithm can provide the best solution to all optimization problems.Based on natural phenomena,species intelligence and their foraging behaviors,scholars have proposed many algorithms based on different theories and technologies,such as traditional evolutionary algorithms represented by genetic algorithms,ant colony optimization,particle swarm optimization,cuckoo search algorithm and firefly algorithm.The Crow search algorithm(CrSA),also inspired by nature,is a new meta-heuristic algorithm which was proposed by Askarzadeh in 2016.The main idea of the algorithm is that crows have the ability to navigate,the wisdom to store and find food,and the smart brain toavoid their food being stolen by other companions.Through the analysis of CrSA,we find that we can use CrSA to solve discrete space search problems,especially to solve feature selection problem.Therefore,Feature Selection Using Crow Search Algorithm(FSCrSA)was proposed.In order to verify the effectiveness of FSCrSA,we used FSCrSA to guide the learning process on three classifiers,namely SVM,J48 and KNN classifiers,and carried out experiments on multiple data sets.Compared with traditional machine learning feature selection algorithms and feature selection algorithm based on evolutionary computing,FSCrSA can select features with strong recognition in the data set,which not only greatly reduces the size of the feature subset,but also improves the classification accuracy.Through the analysis of FSCrSA,we find that the algorithm has limitations in the initialization and search stage,which leads to low dimensional reduction ability.Therefore,we propose the Improved Feature Selection Using Crow Search Algorithm(IFSCrSA).We use the opposition-based learning strategy to set the initial position of the crow in the discrete space.Using the advantages of opposition-based learning,the initial value of the algorithm is not completely random.We choose the better initial value as the initial position to improve the speed of searching.The Lévy flights is used to balance global search and local search.By adjusting the step size of Lévy flights,the algorithm will fall into local optimum prematurely in the early stage,and can quickly converge to global optimum in the later stage;We also propose an update mechanism using greedy strategy to accelerate the convergence rate.The experimental results shows that compared with FSCrSA,IFSCrSA has stronger ability of dimension reduction without reducing classification accuracy.