sklearn分类器-使auc最大化的predict_proba阈值

Question

I have a three-class classification problem. I train the classifier and then plot the ROC for the different classes.

I need to get the threshold for each class which maximizes the TPR and minimizes the FPR. In Matlab, this is returned. Is there a way to retrieve this in python / sklearn?

Thanks.

Answer 1

So my idea looks like this:

import numpy as np
idx = np.linalg.norm(
    (np.array([[0, 1]]) -np.stack([fpr, tpr], axis=1)), 
    axis=1).argmax()
max_thresh = thresholds[idx]

what is done here: stack the FPR and TPR together as a 2D-vector. subtract the upper left corner of each row of the vector and take the norm. This computes the distance. Take the argmax to know in which row the maximum value appears. Finally, return the threshold at that index