A Relational Learner with Hierarchical Background Knowledge

This paper presents a new relational learner that can e ciently handle a large-scale type hierarchy (i.e.,is a relations). Relational Learner with Hierarchical Background Knowledge (RHB) generates typed Prolog programs that discriminate between positive and negative examples on the basis of background knowledge including a large-scale type hierarchy. Previous learners, such as FOIL and GOLEM, virtually fail to learn relations on the basis of a large-scale type hierarchy. This is because a type hierarchy is processed as ordinary background knowledge. RHB provides is a relations with special operations when computing lgg of examples and computing the coding length of typed Prolog programs. RHB, implemented in the LIFE programming language can e ciently learn relations with type hierarchy through LIFE's type handling mechanisms. Experimental results shows that RHB can e ciently handle about 3000 is a relations while still achieving high accuracy.