Adaptive Niching Memetic Algorithm Based Data Clustering

Data clustering is an important task in data mining.Data clustering accomplishes the clustering task by optimizing specific clustering indicator,which is known as an NP-hard problem.Evolutionary algorithm is a global search algorithm which is employed extensively to solve this NP-hard problem.However,the existing evolutionary clustering algorithms are slow and have low precision.To solve this problem,researchers have combined evolutionary algorithms with k-means for data clustering.Although the evolutionary clustering algorithms have achieved good clustering results,there are still some problems that may limit their performance: 1)evolutionary clustering algorithms usually employ fixed-intensity k-means operators;2)it is difficult for the algorithm to effectively maintain the diversity of population,leading to premature convergence;3)evolutionary clustering algorithms need to set the number of clusters in advance.Studying an efficient evolutionary clustering algorithm is still a challenging problem,and this research will also promote the theoretical development of the field of evolutionary computation and clustering.The main work and results of this paper are as follows:1.Aiming at the shortcomings of the use of the k-means operator in the evolutionary clustering algorithm,a generalized k-means usage framework is introduced,which allows the intensity and frequency of k-means to be arbitrarily adjusted during the evolutionary search process.Based on this framework,an adaptive strategy is proposed to dynamically adjust the intensity and frequency of k-means.In addition,in order to prevent the algorithm from falling into local optimum prematurely,an opposite search strategy is proposed.Based on this strategy,an adaptive k-means operator is implemented.Finally,an adaptive and opposite k-meansoperator based memetic algorithm is proposed for data clustering,and good experimental results are obtained.2.Aiming at the the shortcomings of niching methods in maintaining identified local optimal solutions and exploring new local optimal solutions in a single evolutionary search,a multilevel sampling strategy is proposed to drive a large population to store more identified local optimal solutions and explore new local optimal solutions by constructing subpopulation with different search biases.In addition,aiming at the shortcomings of the local search ability of the evolutionary algorithm and the difficulty of providing high-quality local optimal solutions,a crossover-based local search operator is proposed.Combining the multilevel sampling strategy and the crossover-based local search operator,a new differential evolution algorithm is proposed for multimodal optimization problems,and the proposed algorithm is applied to the clustering problem.The experimental results show that the proposed algorithm has strong search ability in multimodal optimization problems and clustering problems.3.Aiming at the disadvantage that the evolutionary clustering algorithms need to set the number of clusters in advance,a neighborhood mutation based differential evolution algorithm is proposed for automatic clustering problems.The proposed neighborhood mutation strategy can promote the search of the optimal cluster number in the early stage of evolutionary search and the search of the optimal clustering partitioning in the end of evolutionary search.In addition,a existing clustering indicator is modified to obtain a new clustering indicator for automatic clustering task.The experimental results show that the proposed algorithm can search for the appropriate number of clusters and the corresponding optimal clustering.