Learning Relational Concepts with Decision Trees

In this paper, we describe two diierent learning tasks for relational structures. When learning a classiier for structures, the rela-tional structures in the training sets are clas-siied as a whole. Contrarily, when learning a context dependent classiier for elementary objects, the elementary objects of the rela-tional structures in the training set are clas-siied. In general, the class of an elementary object will not only depend on its elementary properties, but also on its context, which has to be learned, too. We investigate the question how such classiications can be induced automatically from a given training set containing classiied structures or classiied elementary objects respectively. We present a graph theoretic algorithm that allows the description of the objects in the training set by automatically constructed attributes. This allows us to employ well-known methods of decision tree induction to construct a hypothesis. We present the system INDIGO and compare it with the LINUS-system, known in ILP. The performance of INDIGO is evaluated on the Mesh and the Mutagenicity Data { two datasets that were studied in Machine Learning literature.