As long as I know, for both ROC and PR curves, the classifier performance is usually measured by the AUC. This might indicate that classifiers with equivalent performance might have different ROC/PR curves.
I am interested in the modification of the PR curve shape without a general improvement in performance. Specifically, is there a way to improve the precision at low recalls (say recall lower than 0.25) with a precision reduction in higher recalls?
I have some thoughts about overfitting the model allowing it to learn with high detail the class I what to detect (also requires increasing the detected class's weight). Does it make sense?
