The role of background knowledge : using a problemfrom

Inductive Logic Programming (ILP) systems construct explanations for data in terms of domain-speciic background information. How does the quality of this information aaect the performance of an ILP system? Results from experiments concerned with learning simple programs for list processing suggest that performance is sensitive to the type and amount of background knowledge provided. In particular, background knowledge that contains large amounts of information that is known to be irrelevant to the problem being considered can, and typically does, prevent an ILP system in its search for an correct explanation. The experiments reported here are a completely redone version of those presented in 13]. The results there were withdrawn from publication. 1 In this paper, we examine a more realistic scenario where the relevance of background predicates is graded from being more to less relevant to the problem considered, and no \correct explanation" is known. Theoretical or empirical answers to the performance of an ILP system under such conditions address more directly the concerns of a practitioner. Using a problem from chemistry, this paper reports on an empirical study of the change in the performance of an ILP program as increasingly generic chemical deenitions are added to background knowledge. The problem examined is the task of obtaining rules describing the mutagenic activity of small molecules. The experiments describe the use of the ILP program Progol, and the eeect of adding atom and bond structure, bulk molecular properties, 2 and 3-dimensional descriptions of the molecules as background knowledge. The results show that adding highly relevant background predicates causes signiicant improvements in predictive accuracy, simplicity and time for theory construction. These eeects are less pronounced, although still noticeable, along the scales of accuracy and simplicity as the background knowledge includes increasingly more generic { and hence not as directly relevant { chemical knowledge. For a practitioner concerned with use of ILP, the results suggest that good \\rst-cut" results can be obtained quickly by a simple exclusion of information known to be less relevant.