Active Relational Rule Learning in a Constrained Confidence-Rated Boosting Framework

Boosting is a particularly robust and powerful technique to enhance the prediction accuracy of systems that learn from examples. Although boosting has been extensively studied in the last years for propositional learning systems, only little attention has been paid to boosting in relational learning. The author proposes a successful boosted ILP based relational learning system and an embedded active feature selection technique which together result in a learning time reduction of up to three orders of magnitude compared to state-of-the-art ILP learning systems, while maintaining or even enhancing the interpretability and the predictive accuracy of the induced hypotheses. Unlike existing feature selection methods in relational learning, the feature selection technique used here actively determines feature subsets for learning on the basis of the actual learning process, and avoids the transformation of the given examples into a propositional representation.