Research Of A Clustering Algorithm Based On Undirected Weighed Network

Clustering is a popular data analysis and data mining technique. It means the act ofpartitioning a multi-attribute datasets into homogeneous groups of similar objects. Thegoal of a clustering algorithm is to group sets of objects into classes such that similarobjects are placed in the same cluster while dissimilar objects are in separate clusters.During the process of handling large data, these traditional clustering algorithms havewell known shortcomings such as such as slowness of the convergence, preset classedin large scale data set etc. Combined with the National Natural Science Fund Project"Research on theories of slow and quick combined knowledge extensibleoptimization and sharing in knowledge supernetwork environment"(projectNO.:71071144) and Zhejiang Province Natural Science Fund Project "Research onextension adaptive knowledge service for complex product design"(projectNO.:Z6110334), this paper studies the clustering problem. The clustering of spatialobjects will be transformed to network partition. Clustering Algorithm based onUndirected Weighted Network (CA-UWN) is put forward in this paper.Main work of thesis completion as follows:(1) Introduce the concept of clustering and complex networks, a summary ofprimary clustering methods, clustering algorithm based on complex network andclosely related literature.(2) Propose Clustering Algorithm based on Undirected Weighted Network(CA-UWN). The clustering of spatial objects will be transformed to network partitionin the algorithm. The paper describes the core ideas and thoughts of this algorithm,and its main process includes network building and the segment of the network. Theweighted network for spatial objects is built in term of the similarity. The weightednetwork is adaptively partitioned into groups by exploiting the strength of nodes andthe weight of edges. Rules based on adjacent nodes and Rules based on expansion are proposed to divide network. The paper discusses the meaning of the parameters in thealgorithms and extraction.(3) The proposed clustering algorithm was conducted. Appropriate algorithmparameters are verified by using Iris and Glass in the UCI dataset, the Rand index andthe Jaccard coefficient. The paper assesses the effectiveness of the algorithm, resultsshow that algorithm based on acknowledgement of similarity distribution underlimited number of iterations, can achieve good clustering effect.