Learning probabilistic linear-threshold classifiers via selective sampling