Abstract

Gaussian mixture models (GMMs) are a very useful and widely popular approach for clustering,
but they have several limitations, such as low outliers tolerance and assumption of data normality. Another problem in relation to finite mixture models in general is the inference of an optimal number of mixture components. An excellent approach to solve this problem is model selection, which is the process of choosing the optimal number of mixture components that ensures the best clustering performance. In this thesis, we attempt to tackle both aforementioned issues: we propose using minimum message length (MML) as a model selection criterion for multivariate bounded
asymmetric Student’s t-mixture model (BASMM). In fact, BASMM is chosen as an alternative to improve the GMM’s limitations, as it provides a better fit for the real-world data irregularities. We
formulate the definition of MML and the BASMM, and we test their performance through multiple experiments with different problem settings.
Hidden Markov models (HMMs) are popular methods for continuous sequential data modeling and classification tasks. In such applications, the observation emission densities of the HMM hidden states are typically modeled by elliptically contoured distributions, namely Gaussians or Student’s t-distributions. In this context, this thesis proposes BAMMHMM: a novel HMM with Bounded Asymmetric Student’s t-Mixture Model (BASMM) emissions. This HMM is destined to sufficiently fit skewed and outlier-heavy observations, which are typical in many fields, such as financial or signal processing-related datasets. We demonstrate the improved robustness of our model
by presenting the results of different real-world applications.