0
$\begingroup$

Let's say I have a model which has a binary classification task (Two classes of 0 and 1) and therefore, it outputs a number between 0 and 1, if it is greater than 0.5 we consider it to be class 1 and 0 the other way around.

Now let's say we remove any results in the test set that its output is between two thresholds of 0.4 and 0.6 to make the model more confident. To be more clear, if the output is in that bound, the model just prints "I'm not confident about this image".

Is this approach a good idea in general?

What if the task is about a binary classification of a medical dataset like COVID?

And if so, has this approach used in any recent research?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

In general yes, the predicted probability can be used in this way. However it's important to take into account that this probability is a prediction itself, i.e. the model could be wrong about it. For example the model may predict a probability of 99% positive for an instance which is actually negative. As usual, it cannot be assumed that the model is correct: it has to be evaluated, in particular whether the instances tagged as "not confident" are actually more likely to be wrongly predicted or not.

An important question in this strategy is how to select the bounds of the "not confident" interval, for example arbitrarily choosing [0.4,0.6] may not be optimal.

Improve this answer
$\endgroup$
2
  • $\begingroup$ I suspected what you explained to be the case, thanks for your explanation. My only unanswered question is do you know any research paper that has used this approach? $\endgroup$
    – Amirhossein Rezaei
    Sep 11, 2021 at 10:31
  • $\begingroup$ @AmirhoseinRezaee I know some related work in the NLP area of author identification or author verification, where the question of the confidence of the model in its result is crucial. It might also be remotely related to the task of quality estimation for machine translation, where the goal is to predict how good a predicted translation is. But this problem is very area-specific, I assume that this question of confidence is also studied in medical applications but I'm not knowledgeable about that. $\endgroup$
    – Erwan
    Sep 11, 2021 at 10:47

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.