Abstract：In order to appropriately select the distribution function as the component of the finite mixture model for clustering analysis, a clustering method based on the elliptical contoured mixture model was proposed by theory of generalized multivariate analysis. The finite mixture model was constructed by elliptical contoured distribution family, with similar properties to normal distribution. The label variable was introduced to transform clustering into parameter estimation. The common variable parameters of the model were estimated according to the maximum likelihood and EM algorithm. The function variable parameter was estimated by kernel density estimation theory to obtain iterative update formulas of E and M steps. The elements were classified by the maximum posteriori probability rule. The experiments of different level noise data simulated by uniform distribution show that the proposed method displays good effectiveness and adaptability for non normal data.
2011, Vol. 32 
