Gravar-mail: Estimating classification accuracy in positive-unlabeled learning: characterization and correction strategies