Lazy Propositionalisation for Relational Learning

A number of Inductive Logic Programming (ILP) systems have addressed the problem of learning First Order Logic (FOL) discriminant definitions by first reformulating the FOL learning problem into an attribute-value one and then applying efficient learning techniques dedicated to this simpler formalism. The complexity of such propositionalisation methods is now in the size of the reformulated problem which is exponential when tackling determinate relational problems. We propose a method that selectively propositionalises the FOL training set by interleaving attribute-value reformulation and algebraic resolution. It avoids, as much as possible, the generation of reformulated examples which are not relevant wrt the discrimination task, and still ensures that explicit correct and complete definitions are learned. We present an AQ-like algorithm exploiting this lazy propositionalisation method and then provide a first empirical evaluation on a standard benchmark dataset for ILP, the Mutagenesis problem.