Ph.D. Defense: Jianyu Liu
Ph.D. Thesis Defense
Thursday, April 4th, 2019
324 Hanes Hall
Flexible Graph-Based Learning with Applications to Genetic Data Analysis
(Under the direction of Dr. Yufeng Liu)
With the abundance of increasingly complex and high dimensional data in many scientific disciplines, graphical models have become an extremely useful statistical tool to explore the data structures. In this dissertation, we study graphical models from two perspectives: i) to enhance supervised learning, classification in particular, and ii) graphical model estimation for specific data types. For classification, the optimal classifier is often connected with the feature structure within each class. In the first project, starting from the Gaussian population scenario, we aim to find an approach to utilize the graphical structure information of the features useful for classification. With respect to graphical models, many existing graphical estimation methods have been proposed based on a homogeneous Gaussian population. Due to the Gaussian assumption, these methods may not be suitable for many typical genetic data. For instance, the gene expression data may come from individuals of multiple populations with possibly distinct graphical structures. Another instance would be the single cell RNA-sequencing data, which are featured by substantial sample dependence and zero-inflation. In the second and the third project, we propose multiple graphical model estimation methods for these scenarios respectively. In particular, two dependent count-data graphical models are introduced for the latter case. Both numerical and theoretical studies are performed to demonstrate the effectiveness of these methods.