Great paper, and a tough problem (for one thing, sparse positive samples in training data).