Inference with Aggregation Parfactors: Lifted Elimination with First-Order d-Separation

In this paper we focus on lifted inference for statistical relational models, that is, inference that avoids complete grounding, in models that combine logical and probabilistic assertions. We focus on relational Bayesian networks that can be represented through par factors and aggregation par factors. We present a new elimination rule for lifted variable elimination, and show how to use first-order d-separation to extend the reach of existing elimination rules.