Two Kinds Of Improvement On K-medoids

Data mining is a process to explore the implicit,new,helps decision-making knowledge or rules from mass data,at present in many Internet Co or transaction frequent industries have achieved a wide range of applications,all kinds of results show emerge in an endless stream.Clustering analysis is the most important and the most basic technology in the field of data mining.It is both theoretical and method.K-medoids is a typical unsupervised clustering algorithm based on partition method,a clustering clustering process is simple,high feasibility,advantages of clustering time complexity is close to linear,while the large-scale data mining also showed good support,so it has been developing rapidly in many industries.This article mainly from the evaluation function of the K-medoids clustering algorithm,two improved methods are proposed: the first method is combined with the K-medoids clustering within class scatter matrix and the between class scatter matrix,the aggregate function method,the multi-objective clustering evaluation function is transformed into single objective evaluation function is easy to compute,and gives the K-medoids algorithm of target clustering the evaluation function of the step and flow chart.The second method is to measure according to the compactness and separation of clustering validity index evaluation index of Silhouettes,put forward a K-medoids clustering evaluation measure the compactness and separation function,given the combination step K-medoids clustering algorithm and the flow measurement distance between clusters.According to the two kinds of improved clustering algorithm is proposed in this paper based on the original K-medoids algorithm and PAM(partitioning around medoids(Partitioning Around Medoid,referred to as PAM)comparison,induction and analysis,discusses the similarities and differences of the adaptability between the three methods,and experimental simulation of the three methods in artificial and the actual data set,the accuracy from the clustering results to prove the feasibility of the two algorithms proposed.