ИСТИНА

Block-term decomposition (BTD) was proposed by Lieven de Lathauwer in 2007, and later applied by the research community to different practical problems, to name a few: modelling electroencephalographic measurements of epileptic seizures, blind deconvolution of telecommunication signals, extracting atrial activity from electrocardiograms. In this decomposition input tensor is represented as a sum of structured terms (other tensors). Original BTD model assumes that these terms are consistent - represented in either (constrained) canonical or Tucker format, though examples of using non-consistent formats are also known. In our previous work simultaneous use of terms in different formats was explored in standard machine learning tasks, clustering and classification. This work contributes to extension of block-term decomposition to probabilistic mixing model, namely mixture of probabilistic PCAs with structured mixing matrices.