Markov Logic for Statistical Relational Learning

Опубликовано 12 августа 2016, 2:11

Classical machine learning makes the i.i.d. (independently and identically distributed) assumption on the data instances. Many real world problems have inherent relational structure where the i.i.d. assumption is no longer valid. Representing this relational structure explicitly becomes important for building accurate models of the data. Markov logic is a formalism to achieve the dual goal of representing the relational structure while handling uncertainty using well-founded statistical models. Markov logic models the underlying domain using weighted first-order formulas. These formulas are then compiled into a Markov network with features corresponding to the ground formulas and parameters corresponding to the associated weights. In this talk, I will present the theory behind Markov logic followed by inference and learning techniques. I will also briefly describe an application of Markov logic to the problem of online social network analysis.