Author:

Keywords:

PLL, SRL, statistical, relational, probabilistic, logic, learning, CP-logic

Abstract:

We are confronted with a growing amount of available data which are not only noisy but also have an increasingly complex structure. The field of machine learning, a subfield of artificial intelligence, focuses on algorithms that deduce useful knowledge from data. Our goal is to represent knowledge using probabilistic logic models and to reason with these models in an automated and efficient manner. Such models bring the expressive power of first-order logic to probabilistic models, enabling them to capture both the relational structure and the uncertainty present in such data.In this dissertation we focus on directed probabilistic logic models and more specifically on CP-logic. The aim of CP-logic is to model causal knowledge that explicitly incorporates dynamic concepts such as events and processes. The fundamental building block is the knowledge why events occur and what the effects of these events are. Efficient inference, however, is a bottleneck in CP-logic and in probabilistic logic models in general, affecting also the cost of learning. We have contributed two methods to improve the efficiency of inference and one method for learning.The first method, first-order Bayes ball, extracts the minimal requisite subtheory of a CP-theory necessary to answer a particular query given evidence. Inference becomes more efficient by restricting computations to the minimal requisite subtheory. Contrary to Bayes ball for Bayesian networks, first-order Bayes ball reasons on the first-order level and it returns the requisite part as a first-order CP-theory. The advantages of working on the first-order level are twofold; first, it is more efficient to find the ground requisite network compared to current methods. Second, the resulting requisite network is first-order, permitting it to be used as input for lifted inference methods which exploit the symmetries present in probabilistic logic models to improve the efficiency of inference with several orders of magnitude. Experiments show that first-order Bayes ball improves existing lifted inference methods by reducing the size of the theory that needs to be analyzed and processed.The second method to improve the efficiency of inference is contextual variable elimination with overlapping contexts which capitalizes on deterministic dependencies present in probabilistic logic models. Two special cases of combining deterministic and probabilistic relations are contextual and causal independencies, both commonly used structures in probabilistic models. The original contextual variable elimination technique compactly represents contextual independence by representing the probabilistic model in terms of confactors but cannot handle causal independence because of some restrictions in these confactors. We lift these restrictions and propose a new algorithm to deal with more general confactors. This allows for a more efficient encoding of confactors and a reduction of the computational cost. Experiments show that our algorithm outperforms contextual variable elimination and variable elimination on multiple problems.Lastly, we propose SEM-CP-logic, an algorithm for learning ground CP-logic from data by leveraging Bayesian network learning techniques. To this end, certain modifications are required to parameter and structure learning for Bayesian networks. Most importantly, the refinement operator used by the search must take into account the fine granularity of CP-logic. Experiments in a controlled artificial domain show that learning CP-theories with SEM-CP-logic requires fewer training data than Bayesian network learning.