Structure Learning in Relational Domains

March 27, 2009
Halligan 106
Speaker: Lilyana Mihalkova , University of Texas at Austin
Host: Carla Brodley


The goal of machine learning is to develop techniques that automatically acquire useful knowledge from the wealth of digital data that arises in diverse applications, ranging from social networking and Web search to molecular biology. While the field has progressed remarkably in the past decades, it has focused on learning from data in which each entity is independent of the rest. In contrast, in many practical applications, entities of varying types are connected by a rich set of relations, such as friendship and shared interests in a social network, or chemical bonds in molecular biology.

A fundamental challenge in learning from such relational data is discovering the structure, or the dependencies and regularities present among the relations in the data. Using Markov logic, a general and expressive knowledge representation, I will show how to learn structure accurately and efficiently by transferring a source model that was previously acquired in a different but related domain. I will first describe an algorithm that revises the source model in the case when a significant amount of data from the target domain is available. I will next address transfer learning in the challenging case when target-domain data is severely limited. Finally, I will describe a novel approach for learning structure directly without transfer, and an application of relational learning to resolving ambiguities in Web searches.

Bio: Lilyana Mihalkova is a Ph. D. candidate in the Department of Computer Sciences at the University of Texas at Austin. She joined the department as a recipient of the Microelectronics and Computer Development Fellowship. She is a member of the Machine Learning group, led by Prof. Raymond Mooney. Halfway through her doctorate, she spent the summer of 2007 at Microsoft Research as an intern in the Text Mining, Search, and Navigation group. Her research interests include statistical relational learning, transfer learning, and applications to social networking and web domains.