Research On The Improved Moth-Flame Optimization Algorithm And Its Application For Feature Selection

With the development of technology and the popularization of the Internet,massive data bring new challenges to the research of machine learning and data mining.High-dimensional datasets usually have a lot of irrelevant features,confusing the classification process and significantly increasing the calculation time.How to define useful features and use the relevant features to improve the quality of dataset,become the focus of the current research area.Preprocessing step like feature selection in the field of machine learning and data mining commonly refers to a process of using the candidate algorithm to figure out the optimal feature subset.The main purpose of this step is to select the most useful and relevant features,which directly improves the efficiency of data processing,speeds up the training time and simplifies the classification model to maintain or improve the performance of the classification's accuracy.Feature selection can be regarded as a NP-hard combination optimization problem.The swarm intelligence optimization algorithm is employed to optimize the selection process of feature subsets in this paper.Moth-flame optimization algorithm(MFO)is robust and efficient;alsoeasy to utilize in optimization filed.It has been widely studied to solve some complex optimization problems.Therefore,this paper will primarily focus on the moth flame optimization algorithm and its application on feature selection process.This algorithm has certain advantages compared with other classical swarm intelligence algorithms,but it still has the problems of slow convergence and easy to trap into the local optimum for some complex practical optimization tasks.Therefore,three novel feature selection methods based on the enhanced moth flame optimization algorithm are proposed.The main contributions of this paper are as follows:(1)An enhanced double-strategy evolutionary moth flame optimization algorithm is proposed in the first part to overcome the problem of premature convergence caused by falling into local optimal solutions.First,the Gaussian mutation strategy is introduced into the population updating process to increase the population diversity for the moth flame optimization algorithm.The chaotic local search strategy is employed to the flame updating process of the optimization algorithm for the better capability of escaping from the local optimal.Then,the enhanced MFO algorithm is employed to the kernal extreme learning machine(KELM)model for tuning kernel parameters and selecting the optimal feature subset simultaneously.Finally,the proposed hybrid modelis successfully applied to the prediction of financial optimization problem to investigate its effectiveness.(2)The feature selection task can be regarded as a binary optimization problem.Solutions are limited to two numbers(‘0' and “1”)in the search space.The second part,an improved binary version based on MFO algorithm for feature selection is proposed.The new random mutation operator and the crossover operator are executed in each iteration of the binary optimization algorithm to select the optimal feature subset.The extreme learning machine(ELM)model in this part is chosen to evaluate the feature subset and perform the classification tasks in predicting phenanthrene poisoning to distinguish the most important features in the task.(3)Based on the two improved methods mentioned above,an enhanced MFO algorithm with two types of transfer functions,crossover scheme,ensemble strategy,and the simulated annealing(SA)disturbance mechanism is proposed for feature selection model.The new method provides a way to take advantage of enhanced mechanisms to improve the search efficiency on the feature selection problem and transform solutions into the discrete solutions by two types of transfer functions.Each search individual in this algorithm is assessed by the size of selected features and the error rate of the k-nearest neighbor(KNN).30 benchmark datasets with different dimensions from the UCI repository(University ofCalifornia,Irvine,Repository of machine learning databases)are utilized to assess the search ability and classification performance of the proposed feature selection model.Experimental results show the feasibility of selecting the most informative features for classification purposes and decreasing the classification error rate in the big data environment.