As lazy as it can be

Abstract

Inductive Logic Programming (ILP) is a promising technology for knowledge extraction applications. ILP has produced intelligible solutions for a wide variety of domains where it has been applied. The ILP lack of efficiency is, however, a major impediment for its scalability to applications requiring large amounts of data. In this paper we address important issues that must be solved to make ILP scalable to applications of knowledge extraction in large amounts of data. The issues include: efficiency and storage requirements.We propose and evaluate a set of techniques, globally called lazy evaluation of examples, to improve the efficiency of ILP systems. Lazy evaluation is essentially a way to avoid or postpone the evaluation of the generated hypotheses (coverage tests). To reduce the storage amount a representation schema called interval trees is proposed and evaluated. All the techniques were evaluated using the IndLog ILP system and a set of ILP datasets referenced in the literature. The proposals lead to substantial efficiency improvements and memory savings and are generally applicable to any ILP system.