When seeing the plots of the scores distribution for positive and negative cases, I realized that most approaches tend to produce a peak around one for positive cases, and something relatively flat for negative cases. Can anyone figure out why is this happening? Why don't we see a peak for negative cases and a flat distribution for positive cases? My intuition is that most of us, in order to maximise AUC have used the same (or very similar) number of positive and negative examples in the batches. This could have caused that the ___normality_ of negative cases has not been learnt (too much variability). What do you think should be the strategy to get a peaked distribution for negative examples?