1
$\begingroup$

Suppose we have two binary classifiers based on deep learning. The second classifier is able to tell me with a probability not very high but better than a random guess (let's say 70%), if the prediction of the first classifier is correct (but not which is the correct label).

Therefore, given an unlabeled data set, how could I use the information from the second classifier to retrain the first classifier on the unlabeled data to get a better first classifier?

Improve this question
$\endgroup$
2
  • 2
    $\begingroup$ (1) Are you sure your second classifier is better than random? If one class appears in 70% of cases, and classifier A always classifies as this class, then classifier B can always respond "correct" and would be right 70% of times. (2) What comes to mind is using a third classifier that uses the features, and the outputs from both original classifiers. (3) Have you thought about probabilistic classification? $\endgroup$
    – Stephan Kolassa
    Commented Nov 29, 2022 at 8:58
  • $\begingroup$ (1) Yes, the second classifier is able to predict what the first classifier's performance will be for a given testing set without using the labels. (3) yes, no luck at the moment. (2) This is something I want to try, although I wanted to know if there was a common procedure for this kind of problem. $\endgroup$
    – Daniel Lerch
    Commented Nov 29, 2022 at 10:48

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.