Tushar Khot

Over the past years, Machine Learning (ML) approaches have taken large strides in their predictive accuracy and ease of use, resulting in ML being used in increasing number of domains. At the same time, information has grown exponentially in terms of its size and complexity. Inter-related objects (people, atoms, words, etc.) spread across multiple relations (friends, bonded, dependent, etc.) is now a common occurrence in many domains such as molecular chemistry, medical diagnosis, social networks and information extraction. To deal with noisy multi-relational data, Statistical Relational Learning (SRL) models have been proposed. Unlike most ML approaches that rely on a fixed number of features for every example, SRL models can handle an arbitrary number of features. For instance, a patient can have all their test results, where the number of tests may vary between patients, as features. But due to the increased complexity, SRL models do not scale to large domains, especially when learning the structure of the probabilistic dependencies (e.g., discovering the dependence of parents’ chromosomes on a person’s blood type). My research has mainly concentrated on developing more accurate, scalable structurelearning approaches for SRL models to make them more generally and easily applicable. Since these approaches do not rely on an expert designed model, I was able to use them in diverse domains ranging from natural language processing to medical diagnoses to network analysis.