Zhang and Zhang comment that evaluation bias applied to a data miner changes
how we assess the value of the learned theory. Menzies et.al. report success
with defect predictors that achieve probabilities of detection (pd) and probabilites
of false alarms (pf) of 71% and 23% (respectively). This results seem usefully
high, 
but when they are re-expressed in terms of precision, a different picture emerges
(precisions ranging from