Boosting the Area Under the ROC Curve (Poster M32)
Phil Long (Google) and Rocco Ser vedio (Columbia)
· Ranking problem


 Given examples of members and nonmembers of a class,  output scoring function  Goal: highest scorers mostly members  Standard criterion  Value ½ for random output  Value 1 for perfect output
False positives True positives



·  AUC (Area Under the ROC Cur ve)






·  Theorem


 Can "boost" the AUC (slightly better than ½ implies close to 1)  ... also with misclassification noise at rate  < ½.  Contrasts w/ classification, where 1 ­  best possible [KS05]