ProbLog: a probabilistic programming language for data analysis

invited talkThe ability to reason about large numbers of objects, their attributes, and relationships between them is critical in many applications. Moreover, such applications often require to explicitly reason about uncertainty as well. Examples include finding connections between causes and effects in gene interaction networks, activity recognition and tracking in video, and analyzing information networks that integrate data from sources of different quality. In this talk, I present an overview of our work on ProbLog, which develops both a general methodology and a practical toolbox for data analysis in such uncertain, relational domains. Conceptually, a ProbLog program combines probabilistic choices with a logic program describing the consequences of these choices. For instance, in an information network, probabilistic choices specify the probabilities that the relations corresponding to edges in the network actually hold, whereas the logic program defines additional types of relations that can be derived from the given ones. Based on this probabilistic model, we consider three classes of data analysis tasks. The first is probabilistic inference, that is, computing the probability that a certain relation holds for a specific set of objects, given knowledge on other relations. The second is parameter learning, that is, estimating the probabilities of choices based on observed relational data. The third is structural analysis or structure learning, that is, the task of finding regularities in the data, such as probabilistic association rules, most likely explanations, or subgraph patterns occurring with high probability. More information on ProbLog, including a tutorial and software, is available at http://dtai.cs.kuleuven.be/problog.status: publishe