Experimentation of an expectation maximization algorithm for probabilistic logic programs

Statistical Relational Learning and Probabilistic Inductive Logic Programming are two emerging fields that use representation languages able to combine logic and probability. In the field of Logic Programming, the distribution semantics is one of the prominent approaches for representing uncertainty and underlies many languages such as ICL, PRISM, ProbLog and LPADs. Learning the parameters for such languages requires an Expectation Maximization algorithm since their equivalent Bayesian networks contain hidden variables. EMBLEM (EM over BDDs for probabilistic Logic programs Efficient Mining) is an EM algorithm for languages following the distribution semantics that computes expectations directly on the Binary Decision Diagrams that are built for inference. In this paper we present experiments comparing EMBLEM with LeProbLog, Alchemy, CEM, RIB and LFI-ProbLog on six real world datasets. The results show that EMBLEM is able to solve problems on which the other systems fail and it often achieves significantly higher areas under the Precision Recall and the ROC curves in a similar time.