Context-Specific Independence in Directed Relational Probabilistic Models and its Influence on the Efficiency of Gibbs Sampling

There is currently a large interest in relational probabilistic models. While the concept of context-specific independence (CSI) has been well-studied for models such as Bayesian networks, this is not the case for relational probabilistic models. In this paper we show that directed relational probabilistic models often exhibit CSI by identifying three different sources of CSI in such models (two of which are inherent to the relational character of the models). It is known that CSI can be used to speed up probabilistic inference. In this paper we show how to do this in a general way for approximate inference based on Gibbs sampling. We perform experiments on real-world data to analyze the influence of the three different types of CSI. The results show that exploiting CSI yields speedups of up to an order of magnitude.