Author:

Abstract:

Statistical relational learning formalisms combine first-order logic with probability theory in order to obtain expressive models that capture both complex relational structure and uncertainty. Despite the significant progress made in this field, several important challenges remain open. First, the expressivity of statistical relational learning comes at the cost of inefficient learning and inference in large-scale problems that contain many objects. Second, while many real-world relational domains are hybrid in that they contain objects that are described by both continuous and discrete properties, little attention has been paid to learning from such data. Third, most formalisms ignore the dynamic nature of real-world problems by considering only the static aspects captured by a single snapshot of time in the dynamic process. This thesis tries to tackle these shortcomings and makes the following four contributions. First, we propose a graph-sampling based approach that approximately counts the number of pattern occurrences in the data, which enables scaling up parameter learning of statistical relational models. Second, we propose a novel statistical relational learning formalism that models hybrid relational domains. Third, we designed the first structure learning algorithm that is able to learn hybrid relational models. Fourth, we adapted our algorithm to learn temporal dependencies present in the data. We demonstrate the utility of our approaches on several challenging applications, such as planning in a real-world robotics setup, and learning from financial and citation data.