PhD Oral Exam - Kamal Maanicshah Mathin Henry, 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

Unsupervised learning has been an interesting area of research in recent years. Novel algorithms are being built on the basis of unsupervised learning methodologies to solve many real world problems. Topic modelling is one such fascinating methodology that identifies patterns as topics within data. Introduction of latent Dirichlet Allocation (LDA) has bolstered research on topic modelling approaches with modifications specific to the application. However, the basic assumption of a Dirichlet prior in LDA for topic proportions, might not be applicable in certain real world scenarios.

Hence, in this thesis we explore the use of generalized Dirichlet (GD) and Beta-Liouville (BL) as alternative priors for topic proportions. In addition, we assume a mixture of distributions over topic proportions which provides better fit to the data. In order to accommodate application of the resulting models to real-time streaming data, we also provide an online learning solution for the models. A supervised version of the learning framework is also provided and is shown to be advantageous when labelled data are available.

There is a slight chance that the topics thus derived may not be that accurate. In order to alleviate this problem, we integrate an interactive approach which uses inputs from the user to improve the quality of identified topics. We have also tweaked our models to be applied for interesting applications such as parallel topics extraction from multilingual texts and content based recommendation systems proving the adaptability of our proposed models. In the case of multilingual topic extraction, we use global topic proportions sampled from a Dirichlet process (DP) to tackle the problem and in the case of recommendation systems, we use the co-occurrences of words to our advantage.

For inference, we use a variational approach which makes computation of variational solutions easier. The applications we validated our models with, show the efficiency of proposed models.