Temporal and Relational Graphical Models for Causality: Representation and Learning

28 Jul
Add to Calendar
Friday, 07/28/2017 10:00am to 12:00pm
Computer Science Building, Room 142
Ph.D. Thesis Defense

"Temporal and Relational Graphical Models for Causality: Representation and Learning"

Discovering causal dependence is central to understanding the behavior of complex systems and to selecting actions that will achieve particular outcomes. The majority of work in the area of causality has focused on propositional domains, where data instances are considered to be independent and identically distributed (i.i.d.). However, many real-world domains are inherently relational, i.e., they consist of multiple types of entities that interact with each other, and temporal, i.e., they change over time. This thesis focuses on causal modeling for these more expressive relational and temporal domains. The contributions of this thesis are geared towards an in-depth investigation of the properties of relational models and towards extending their expressivity to include a temporal dimension. Specifically, we first investigate alternative ways to ground relational models, and we provide an in-depth analysis of the impact of alternative grounding semantics for feature construction, causal effect estimation, and model selection. Then, we extend relational models to represent discrete time. We generalize the theory of d-separation for this class of temporal and relational models. Finally, we provide a constraint-based algorithm, TRCD, to learn the structure of temporal relational models from data.

Advisor: David Jensen