SRL 2003 IJCAI 2003 Workshop on Learning Statistical Models from Relational

We present a general approach to speeding up afamily of multi-relational data mining algorithmsthat construct and use selection graphs to obtain theinformation needed for building predictive mod-els (e.g., decision tree classifiers) from relationaldatabase. Preliminary results of our experimentssuggest that the proposed method can yield 1-2 or-ders of magnitude reductions in the running timeof such algorithms without any deterioration inthe quality of results. The proposed modificationsenhance the applicability of multi-relational datamining algorithms to significantly larger relationaldatabases that would otherwise be not feasible inpractice.