Boosting tuple propagation in multi-relational classification

Multi-relational classification is a mining method aiming at building classifiers for the tuples in some target relation based on its own data as well as on the data possibly dispersed over other non-target relations, by exploiting the relationships among them formalized via foreign key constraints. While improving on the efficacy of the resulting classifiers, propagating data via the foreign key constraints deteriorates the scalability of the underlying algorithm. In the paper, various techniques are discussed to efficiently implement this propagation task, and hence to boost performances of current multi-relational classification algorithms. These techniques are based on suitable adaptations of state-of-the-art query optimization methods, and are conceived to be coupled with database management systems. A system prototype integrating all the techniques is illustrated, and results of experimental activity conducted on top of it are eventually discussed.