Caraf: Complex Aggregates Within Random Forests

This paper presents an approach integrating complex aggregate features into a relational random forest learner to address relational data mining tasks. CARAF, for Complex Aggregates within RAndom Forests, has two goals. Firstly, it aims at avoiding exhaustive exploration of the large feature space induced by the use of complex aggregates. Its second purpose is to reduce the overfitting introduced by the expressivity of complex aggregates in the context of a single decision tree. CARAF compares well on real-world datasets to both random forests based on the propositionalization method RELAGGS, and the relational random forest learner FORF. CARAF allows to perform complex aggregate feature selection.