EcoSta 2018: Registration
View Submission - EcoSta2018
A0594
Title: Assessment of model fit in clustering via mixture models Authors:  Suren Rathnayake - The University of Queensland (Australia) [presenting]
Abstract: Some issues are considered associated with the fitting of a normal mixture model to cluster data known to be drawn from an unknown number of distinct classes. The use of the likelihood ratio statistic is investigated for tests on the smallest number of components in the mixture model for it to be compatible with the observed data. Also, under the implicit assumption that the clusters implied by the fitted mixture model are in correct correspondence with the external existing classes, we investigate further the estimation of the accuracy of the implied clustering. For this purpose, an estimator of the overall correct allocation rate is formed by averaging the maximum of the (estimated) posterior probabilities of component membership for each observation.