Research On K-medoids Clustering Algorithm Based On Improved Granular Computing

With the rapid development of database storage technology, massive data is piled up which is a headache for organizations. In order to extract useful information from massive data, data mining comes into being. As one of important research branches in data mining, cluster analysis is aim to divide the data into meaningful or useful clusters by analyzing the similarity among the data.This paper discusses the K-medoids clustering algorithm and proposes improved strategies. These strategies include the use of improved granular computing to initialize the medoids, propose three different mediods search update strategies and improve fitness function. The main research work of this paper shows as follows:To solve the problem that traditional K-medoids clustering algorithm is sensitive to the initial selection of the mediods, a novel algorithm initialize K medoids based on improved granular computing. The results of experiments show that this novel algorithm tested on standard data set Iris in UCI, generating K initialized medoids which are located in different clusters respectively, as a result, effectively avoid the traditional clustering algorithm initialize sensitive issue.To solve the problems that slow convergent speed and poor accuracy of traditional K-medoids clustering algorithm, under the premise of effective initialization, a granule iterative search strategy is proposed in order to reduce the number of iterations in the processes of clustering algorithm. Meanwhile, a novel fitness function balanced inner-cluster distance and among-clusters distance is proposed in order to enhance the adaptability and precision of clustering algorithm.To solve the problems that the search blindness and not high accuracy of traditional K-medoids clustering algorithm, under the premise of effective initialization too, two kinds of new medoids search strategies are proposed--nearby search strategy and cluster search strategy in order to avoid the traditional K-medoids cluster algorithm’s defect of blind global search. What’s more, a novel criterion function is proposed in order to enhance the precision of clustering algorithm.