View article

This paper presents a new approach to constructive induction, Discrimination-Based Constructive induction(DBC), which invents useful predicates in learning relations. Triggered by failure of selective induction, DBC finds a minimal set of variables forming a new predicate that discriminates between positive and negative examples, and induces a definition of the invented predicate. If necessary, it also induces subpredicates for the definition. Experimental results show that DBC learns meaningful predicates without any interactive guidance.