Abstract

This thesis proposes innovative clustering approaches using finite and infinite mixture models to analyze medical data and human activity recognition.
These models leverage the flexibility of a novel distribution, the multivariate McDonald’s Beta distribution, offering superior capability to model data of varying shapes. We introduce a finite McDonald’s Beta Mixture Model (McDBMM), demonstrating its superior performance in handling bounded and asymmetric data distributions compared to traditional Gaussian mixture models.
Further, we employ deterministic learning methods such as maximum likelihood via the expectation maximization approach and also a Bayesian framework, in which we integrate feature selection. This integration enhances the efficiency and accuracy of our models, offering a compelling solution for real-world applications where manual annotation of large data volumes is not feasible.
To address the prevalent challenge in clustering regarding the determination of mixture components number, we extend our finite mixture model to an infinite model. By adopting a nonparametric Bayesian technique, we can effectively capture the underlying data distribution with an unknown number of mixture components.
Across all stages, our models are evaluated on various medical applications, consistently demonstrating superior performance over traditional alternatives.
The results of this research underline the potential of the McDonald’s Beta distribution and the proposed mixture models in transforming medical data into actionable knowledge, aiding clinicians in making more precise decisions and improving health care industry.