Structure Learning with Hidden Data in Relational Domains

Recent years have seen a surge of interest in learning the structure of Statistical Rela-tional Learning (SRL) models, which com-bine logic with probabilities. Most of these models apply the closed-world assumption i.e., whatever is not observed is false in the world. We consider the problem of learn-ing the structure of SRL models in the pres-ence of hidden data, i.e. we open the closed-world assumption. We develop a functional-gradient boosting algorithm based on EM to learn the structure and parameters of the models simultaneously and apply it to learn two kinds of models – Relational Dependency Networks (RDNs) and Markov Logic Net-works (MLNs). Our results in two testbeds demonstrate that the algorithms can effec-tively learn with missing data. 1