PhD Oral Exam - Basim Tareq M Alghabashi, Information and Systems Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

The exponential growth of heterogeneous data, including text, images, audio, video, and sensor measurements, has created an increasing demand for intelligent data analysis techniques capable of extracting meaningful knowledge from large-scale and high-dimensional datasets. Among machine learning approaches, unsupervised learning has attracted considerable attention due to its ability to discover hidden structures from unlabeled data. In particular, finite mixture models (FMMs) have become an effective statistical framework for data modeling, clustering, classification, and segmentation across numerous application domains. However, their performance strongly depends on the choice of the underlying probability distribution, accurate parameter estimation, automatic model selection, and the identification of relevant features.

This thesis addresses these challenges by proposing a unified statistical learning framework based on the Generalized Gamma Mixture Models (GGMMs) for modeling multi-dimensional positive-valued data. First, a novel multi-dimensional GGMM is developed, where the model parameters are estimated using the Maximum Likelihood (ML) criterion within the Expectation-Maximization (EM) framework. Since no closed-form solution exists for one of the model parameters, the Newton–Raphson algorithm is incorporated into the estimation procedure. Next, the proposed framework is extended by integrating the Minimum Message Length (MML) criterion to automatically determine the optimal number of mixture components. To further improve clustering performance, a feature selection mechanism is incorporated to simultaneously estimate feature relevancy and model parameters, thereby reducing the influence of irrelevant features while improving clustering accuracy.

The proposed methods are extensively evaluated using both synthetic and real-world datasets, including texture, scene, and shape image clustering problems. Finally, the developed framework is applied to a challenging transportation application, namely road condition monitoring. The proposed GGMM is employed for automatic pothole detection and mask generation, where the generated masks are subsequently utilized as training labels for a U-Net deep learning model to perform pothole segmentation without manual annotation. In addition, an intelligent recommendation system is developed to prioritize maintenance actions according to the detected pothole severity. Experimental results consistently demonstrate that the proposed GGMM framework outperforms conventional Gaussian and Gamma mixture models while providing an effective, interpretable, and scalable solution for multi-dimensional data modeling and intelligent transportation applications.