In T128087, @Sabya notes that we generally get better ROC-AUC when we drop the weight balancing strategy. What effect would this have on the outputs of our models if we did this across the board?

Recreate this graph for two models trained with and without balancing and compare.