Trading Expressivity for Efficiency in Statistical Relational Learning

Statistical relational learning (SRL) combines state-of-the-art statistical modeling with relational representations. It thereby promises to provide effective machine learning techniques for domains that cannot adequately be described using a propositional representation. Driven by new applications in which data is structured, interrelated, and heterogeneous, this area of machine learning has recently received increasing attention. However, combining statistical modeling and relational representations also poses new challenges. There is a trade-off between the expressivity of a machine learning formalism and its computational efficiency, as a higher expressivity entails a larger search space during learning. Propositional machine learning techniques are at one end of this trade-off, while approaches that combine the full power of statistical and relational learning are at the other end. This thesis presents a collection of simple SRL techniques that focus on computational efficiency rather than maximum expressivity, and thereby occupy an intermediate position in the outlined expressivity-efficiency trade-off. The thesis has three main contributions. We first introduce dynamic propositionalization approaches, which provide a simple but principled integration of relational and statistical learners. Dynamic propositionalization is shown to outperform more traditional static propositionalization approaches, while maintaining computational efficiency. A second part presentsMarkov models for relational sequences, where sequence elements can be logical atoms or complete logical interpretations. By restricting attention to fully observable data and employing a Markov assumption, inference and learning in the resulting formalisms is significantly easier than in more general SRL systems. In a final part, we present two structured probabilistic models that are tailored to particular application domains, namely haplotype reconstruction and activity recognition. These two domains could be modeled using general-purpose statistical relational sequence models; however, the restriction to a particular domain again allows us to derive more efficient special-purpose inference and learning algorithms. The approaches presented throughout the thesis are evaluated in several relational realworld domains, including structure-activity prediction for chemical compounds, web page classification, modeling user behavior in mobile phone networks, and modeling massively multiplayer online games.