A0471
Title: Matrix-variate hidden Markov models for robust clustering and anomaly detection
Authors: Salvatore Daniele Tomarchio - University of Catania (Italy) [presenting]
Abstract: The matrix-variate hidden Markov model (HMM) framework is extended by introducing two novel model families that employ the matrix-variate t distribution and a contaminated normal distribution. These extensions enhance the modeling of heavy tails, improve clustering performance, and support the detection of atypical matrices. Parameter estimation relies on two tailored expectation-conditional maximization (ECM) algorithms, both of which are implemented in the MatrixHMM R package. Simulation studies demonstrate the models' accuracy, robustness, and effectiveness in outlier identification. Finally, an application to real data illustrates how the models can be used to explore labor market trends across Italian provinces.
