The blog Recidivism, and the Failure of AUC published on Statistics.com showed how the use of “Area Under the Curve” (AUC) concealed bias against African-Americans defendants in a Model predicting recidivism, that is, which defendants would re-offend. There, a model varied greatly in its performance characteristics depending on whether the defendant was white or black. Though both situations resulted in virtually identical AUC measures, they led to very different false alarm vs. false dismissal rates. So, AUC failed the analysts relying on it, as it quantified the wrong property of the models and thus missed their vital real-world implications.
This post first appeared on Elder Research Data Science & Machine Learning Blog, please read the originial post: here