Biases and Their Effects in Inductive Logic Programming

In inductive learning, the shift of the representation language of the hypotheses from attribute-value languages to Horn clause logic used by Inductive Logic Programming systems accounts for a very complex hypothesis space. In order to reduce this complexity, most of these systems use biases. In this paper, we study the innuence of these biases on the size of the hypothesis space. For this comparison, we rst identify the basic constituents the biases are combined of. Then, we discuss the restrictions set on the distribution of terms in the clause by the constituents of the bias. The eeects of several constituents and of some combinations are shown by seven experiments.