A1016
Title: Functional projection Gaussian mixture
Authors: Stefano Antonio Gattone - University G. d'Annunzio of Chieti-Pescara (Italy) [presenting]
Roberto Rocci - Sapienza University of Rome (Italy)
Abstract: The purpose is to introduce a novel model-based approach that simultaneously performs clustering and dimension reduction of functional data. The method assumes that observed functional data are distributed as a finite mixture of Gaussian processes, with differences among components represented in terms of means and covariances within a reduced-dimensional functional subspace. The inference is drawn conditionally at the points where the curves are evaluated, following a penalized maximum likelihood approach to ensure smooth estimates of the centroids. An EM-type algorithm is presented to compute these estimates. The effectiveness of this simultaneous clustering and dimension reduction method is demonstrated through applications to both real and simulated data. This approach offers a powerful alternative to traditional tandem analysis, which typically involves first applying dimension-reduction techniques to the data and then performing clustering on the reduced space. However, tandem analysis may not be optimal because the reduced space identified in the first step might not contain the features most relevant for clustering, potentially leading to suboptimal clustering results.
