DATA SEGMENTATION USING MIXTURE REGRESSION MODELS WITH GENERALIZED GAUSSIAN DISTRIBUTION AND K-MEANS

Data segmentation using mixture regression models gained lot of momentum due to their ready applicability in market analytics, business analytics, financial analytics, supply chain analytics , Human Resource analytics etc. In regression analysis it is customary to assume that error term follows a Gaussian distribution. Gaussian distribution has several drawbacks such as being mesokurtic and the model may not serve well for all types of data. Hence, in this paper we develop data segmentation method using mixture of regression models with Generalized Gaussian Distributed (GGD) errors. The GGD includes leptokurtic, platykurtic and Gaussian distribution as particular cases. The model parameters are estimated using Expectation Maximization (EM) algorithm. The initialization of the parameters is done by using k-means algorithm. The data segmentation algorithm is derived using component maximum likelihood under Bayesian framework. The utility of the proposed algorithm is demonstrated with market segmentation. The performance of the algorithm is evaluated by computing segmentation performance metrics such as accuracy, misclassification rate, precision. It is observed that this method performs much better than the earlier data segmentation methods having Gaussian distributed errors for the data sets having leptokurtic and platykurtic response variables.