On Multi-class Problems and Discretization in Inductive Logic Programming

In practical applications of machine learning and knowledge discovery, handling multi-class problems and real numbers are important issues. While attribute-value learners address these problems as a rule, very few ILP systems do so. The few ILP systems that handle real numbers mostly do so by trying out all real values applicable, thus running into eeciency or overrtting problems. The ILP learner ICL (Inductive Constraint Logic), learns rst order logic formulae from positive and negative examples. The main characteristic of ICL is its view on examples, which are seen as interpretations which are true or false for the target theory. The paper reports on the extensions of ICL to tackle multi-class problems and real numbers. We also discuss some issues on learning CNF formulae versus DNF formulae related to these extensions. Finally, we present experiments in the practical domains of predicting mutagenesis, nite element mesh design and predicting biodegradability of chemical compounds.