Penalized model-based clustering for three-way data structures

Recently, there has been an increasing interest in developing statistical methods able to find groups in matrix-valued data. To this extent, matrix Gaussian mixture models (MGMM) provide a natural extension to the popular model-based clustering based on Normal mixtures. Unfortunately, the overparametrization issue, already affecting the vector-variate framework, is further exacerbated when it comes to MGMM, since the number of parameters scales quadratically with both row and column dimensions. In order to overcome this limitation, the present paper introduces a sparse model-based clustering approach for three-way data structures. By means of penalized estimation, our methodology shrinks the estimates towards zero, achieving more stable and parsimonious clustering in high dimensional scenarios. An application to satellite images underlines the benefits of the proposed method.