I have a CNN model for an imbalanced image classification problem. I’m experimenting with a theory that is supposed to improve the accuracy of the model. Since I’m dealing with imbalanced data, I’m reporting f1, prec/rec auc and the average of recalls for each class (mean of the confusion matrix diagonal values). It seems like the new modification only improves PRAUC but f1 and average recall still stays comparable to the baseline. How can I interpret this?