calibration of predicted probabilities

[tholor]

Good ROC curves per subject do not necessarily lead to an good "overall ROC Curve".
Calibration of the predicted probabilities between the subjects is likely to result in a higher score!

http://www.kaggle.com/c/seizure-prediction/forums/t/10383/leaderboard-metric-roc-auc
"You could plot ROC curves for each patient on your cross-validation predictions and then adjust the prediction values so that the curves 'line up' in such a way to optimise for maximum global AUC. Then use those to patch your final predictions before submission... but there's no guarantee that your predictions on the test segments will produce similar ROC curves for this to be effective. "