Research On Finite Gaussian Mixture Model Cluster

Cluster analysis always play an important role in statistics, with therapid development of information technology, cluster analysis becomemore and more important in data processing and data analysis. In recentyears, Model-based clustering algorithm arise a lot of attention in clusteralgorithm research. Its main idea is: assuming the data to obey a mixturedistribution, each cluster can be matched with a branch model, and thendetermine the data belongs to which cluster. Gaussian mixture model iswidely used in the model cluster. When the order number of the modelgreater than1and is a limited, it is a finite Gaussian mixture model. Inthe process of using Gaussian mixture model clustering, EM Algorithm isa kind of effective method in its maximum likelihood estimation. EMalgorithm is a kind of maximum likelihood estimate method, which needto constantly iteration and then estimate the values of parameters. Italways been used in adding data. The core thought of the EM Algorithmis using existing prior knowledge and then iteration likelihood function,let it converges to a certain optimal value, EM algorithm can greatlysimplify the process of parameter estimation in finite Gaussian mixturemodel clustering algorithm.This paper summarizes the current research of based on the Gaussianmixture model clustering algorithm, research its cluster principle and theparameter estimation of EM algorithm in the process of cluster. Based on the current research, we propose the corresponding solution method.First, proposing a initialization method that simple, effective and lessamount of calculation based on the Tri-sectional. Second, proposing theBootstrap-EM algorithm，avoid EM algorithm falling into local optimumparameter estimation. Third, combing EM algorithm and PCA, proposingthe PCA-EM algorithm, it can effectively improve the accuracy and thespeed. While also can solve the problem that EM algorithm can notiteration when the covariance matrix is singular. Last, writing theprogram in R, testing these three methods by standard data set in UCI.The results show that the several algorithms that proposed can effectivelysolve the problem in Finite Gaussian mixture model cluster, and improvethe accuracy of clustering results.