PhD Oral Exam - Fares Alkhawja, 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

In the era of digitization, the immense amounts of data generated on a daily basis, from abundant media streams across various modalities, make it challenging to infer patterns that can further help industries draw conclusions or make decisions. These challenges arise from the unstructured nature of such data, its high dimensionality, and the complex correlations among its features. As a result, traditional analytical methods tend to struggle to capture the underlying structure of these data sources and fail to scale effectively. Moreover, existing deep learning models not only require labeled data for training, but also suffer from decision processes that are neither traceable nor explainable. Statistical generative models have witnessed a notable adoption across many applications recently due to their unsupervised nature, efficient performance under data scarcity, and interpretable coefficients that facilitate explainability. The Dirichlet distribution is widely used in generative models, including mixture and topic models. Despite its flexibility, it still has some limitations, namely its restrictive covariance matrix and its direct proportionality between its mean and variance. These limitations affect its ability to generalize to data with more complex correlation structures and to capture latent patterns in high-dimensional, unstructured data.

To address these limitations, this thesis demonstrates a generalization of the Dirichlet distribution, namely the Nested Dirichlet distribution (NDD), in the context of generative modeling frameworks, due to its greater flexibility and hierarchical structure. Moreover, the NDD generalizes the well-known generalized Dirichlet distribution (GDD) by allowing a more general covariance matrix and relaxing the dual-branch constraint imposed by the GDD, resulting in a more flexible tree structure. This enables the NDD to better handle complex sparse data that suffers from the curse of dimensionality. Furthermore, its hierarchical structure promotes the adoption of NDD in hierarchical feature learning and classification frameworks. This thesis demonstrates the power and superiority of the NDD across multiple works. First, it presents the enhanced performance of the NDD over the Dirichlet distribution and the GDD in mixture and hierarchical Bayesian models. This is achieved by introducing beneficial extensions that reduce model complexity and improve the use of the NDD's hierarchical structure. Second, it introduces hierarchical topic models based on the NDD to inherit its hierarchical structure in document--topic or topic--word representations, within topic models such as the Latent Dirichlet Allocation. This includes parametric and non-parametric hierarchical topic models, in addition to incorporating domain knowledge as a topic--word prior through a Nested Dirichlet Forest. Finally, it employs an NDD-based hierarchical topic model in non-intrusive load monitoring (NILM) to disaggregate appliance energy consumption in residential houses by augmenting the model with point processes. To account for the inherently event-driven and temporally correlated nature of appliance usage, temporal correlations among appliances are captured using the Hawkes process and the Transformer Hawkes process. The results reported in this thesis highlight the advantages of the proposed NDD-based frameworks and pave the way for further improvements in future work.