Abstract

Statistical relational learning (SRL) is a subfield of machine learning that combines relational representations (from databases and AI) with probabilistic modeling techniques (most often graphical models)for modeling network data (typically richly structured multi-relational and multi-model networks).  In this talk, I will briefly review some SRL modeling techniques, and then I will introduce hinge-loss Markov random fields (HL-MRFs), a new kind of probabilistic graphical model that supports scalable collective inference from richly structured data.  HL-MRFs unify three different approaches to convex inference: LP approximations for randomized algorithms, local relaxations for probabilistic graphical models, and inference in soft logic.  I will show that all three lead to the same inference objective.  This makes inference in HL-MRFs highly scalable.   Along the way, I will describe several successful applications of HL-MRFs and I will describe probabilistic soft logic, a declarative language for defining HL-MRFS.