Just skimmed the abstract — on a practical level — can this be used to better train a global ranking function given subsets of example ranked data?