CUD@ILP: a GPU-based Massively Scalable Inductive Learning Algorithm

Inductive Logic Programming developed within Logic Programming in the Nineties; nowadays several ILP frameworks capable of learning monotonic and non monotonic programs from hypotheses and examples exist. In this paper we focus on FOLD-RM, an inductive learning algorithm capable of learning hypotheses for multiclassification problems without the need of using any additional ASP solver. We describe how such an algorithm can be effectively scaled to deal with real world datasets by means of a massive parallel implementation. We show that, by exploiting the massive computational power offered by modern GPUs, our parallel implementation CUD@ILP significantly outperforms FOLD-RM in all the tests considered.