Application of Graph-Based Concept Learning to the Predictive Toxicology Domain
Jesús A. González, Lawrence B. Holder, Diane J. Cook · 2001
Introduction For concept learning systems, data representation is crucial. A good representation might make possible the learning of a concept that was not learnable using other representations. For example, in the case of the earthquake domain (earthquake data such as its location, epicenter, intensity, depth, type, etc.) we can just use an attribute-value representation. However, we are missing important information such as the distance between the earthquakes' epicenters and the difference in time between the earthquakes. This information is lost in an attribute-value representation; information that could make possible the learning of an important concept. In the case of ILP systems, data is represented in First-Order Predicate Calculus (FOPC) in the form of Prolog logic programs. This representation supports structural relations. ILP systems have been successful in structural domains: Progol in the Chemical Carcinogenicity domain [8] and FOIL for learning patterns in Hypertext d