Efficient Learning of Statistical Relational Models

Machine Learning has been successfully applied to many prediction problems in varying domains. But standard techniques assume that the examples are independent of each other and have the same number of features. In many domains, the objects can be inter-related and have different number of features. To build probabilistic models over such data, Statistical Relational Learning (SRL) methods have been proposed, which combine first-order logic representation with probabilities. But due to their high expressivity, learning the structure of SRL models can be computationally intensive.;I present a structure-learning approach that learns multiple weak rules of thumb via functional-gradient boosting. My approach can be used to learn the structure of two popular SRL models. I empirically demonstrate it to be more accurate and computationally faster than state-of-the-art methods.;To further increase the applicability of my approach, I extend it to handle missing data by deriving an Expectation-Maximization approach for relational models. To handle Natural Language Processing domains with only positive labeled examples, I present and evaluate a non-parametric approach for relational one-class classification using a tree-based relational distance measure. Apart from learning models, this thesis also explores knowledge representation in Markov Logic Networks (MLN). I present and evaluate an approach that can convert multi-level combination functions along with their corresponding parameters into MLN clauses. I present an algorithm for converting two combination functions into MLNs and show the correctness of my transformation.;Finally this thesis shows how my approach can be used for Alzhiemer's disease prediction from MRI images as well as to augment expert rules for temporal relation extraction. I present my approach for a large-scale novel relation extraction task, where I process terabytes of streaming data to detect changes in extracted relations.;Overall, this thesis presents multiple structure-learning approaches for SRL, starting from a boosting-based algorithm, which is extended to handle missing values via EM. Next, I present a structure-learning approach for one-class classification by learning a relational distance metric. I present application of these structure-learning approach on multiple SRL datasets and real-world tasks.