Title: probabilistic models with predicates and elements -- recent results

Speaker: Eyal Amir

Abstract: Probabilistic models are traditionally specified at a
propositional level, i.e. as distributions over possible assignments
to random variables. Work in AI and Knowledge Representation over the
past 20 years resulted in specification languages involving predicates
and a universe of elements for such models. Recent results show that
inference approaches for such models are more tractable than those for
traditional representations. In this talk I will present one such
relational-probabilistic specification language and will discuss
recent results about inference with that language. I will also present
applications of those to (a) link prediction in social networks and
(b) econometric models with hidden variables.