Proxy for AUC When Training Neural Network?

nok · February 24, 2018, 6:00pm

What metrics is a good proxy for training deep learning model with an objective to maximize AUC? Is cross_entropy the best choice?

ecdrid · February 25, 2018, 2:44am

We can use roc_auc_score itself from sklearn.metrics

Don’t use accuracy…(that’s all I know…)

Have a look at this link

nok · February 25, 2018, 7:59am

I know about sklearn but we cannot use auc as a training metric right? It’s not differentiable and how would we backprop it.

ecdrid · February 25, 2018, 8:49am

Didn’t see that you will use deep learning

machinethink · February 25, 2018, 9:59am

I don’t know if it’s appropriate in all situations, but you can try the Dice loss. This loss function is 1 - Dice coefficient, and the Dice coefficient is also known as the F1-score, which combines precision and recall into one number. It’s not exactly the same as the AUC but both try to maximize precision and recall.