A0442
Title: Model-based clustering for tensor-variate data
Authors: Salvatore Daniele Tomarchio - University of Catania (Italy) [presenting]
Antonio Punzo - University of Catania (Italy)
Luca Bagnato - Catholic University of the Sacred Heart (Italy)
Abstract: More flexible statistical methodologies are necessary with the increasing complexity of real data. One type of data that exemplifies this need is tensor-variate (or multi-way) structures. However, real data often includes atypical observations that render the traditional normality assumption unsuitable. Two novel tensor-variate distributions that are heavy-tailed generalizations of the tensor-variate normal distribution are introduced to address this issue. These distributions are then used to construct finite mixture models for model-based clustering. The eigendecomposition of the components' scale matrices is utilized to reduce complexity in the models, resulting in two families of parsimonious tensor-variate mixture models. The parameter estimation employs variations of the EM algorithm. Since the number of parsimonious models is dependent on the order of the tensors, strategies are implemented to shorten the initialization and fitting processes. The effectiveness of these procedures is evaluated through simulated data analyses. Real data are also investigated.
