HTILDE

Abstract

Nowadays, many organizations have relational databases with millions of records and an important question is how to extract information from them. This work proposes HTILDE (Hoeffding TILDE) to handle very large relational databases, based on the Inductive Logic Programming (ILP) system TILDE (Top-down Induction of Logical Decision Trees) and the propositional Very Fast Decision Tree (VFDT) learner. It is an incremental and anytime algorithm that uses the Hoeffding bound to find out the amount of examples that must be considered for choosing the best test for a node. The results show that, compared to TILDE, HTILDE generates theories from very large relational datasets more efficiently without harming their quality measures (F-measure, precision, recall and accuracy). Also, HTILDE learns less complex theories than TILDE.