I am working on a problem where the labels/response variables take the form of #successes / #attempts. Clearly the goodness of the label depends on the number of attempt so I'd like to avoid the model to learn corner cases like y=0, y=1 that essentially occur because not enough attempts have been made.
We generally frame this problem as either a regression task with mse loss and weights given by #attempts or by looking at it as a classification task with label in [0, 1] and weights equal to #attempts - #successes and #successes respectively and trained through binary cross entropy.
Do you have any paper to recommend that tackle this problem?
thanks in advance
I am working on a problem where the labels/response variables take the form of #successes / #attempts. Clearly the goodness of the label depends on the number of attempt so I'd like to avoid the model to learn corner cases like y=0, y=1 that essentially occur because not enough attempts have been made.
We generally frame this problem as either a regression task with mse loss and weights given by #attempts or by looking at it as a classification task with label in [0, 1] and weights equal to #attempts - #successes and #successes respectively and trained through binary cross entropy.
Do you have any paper to recommend that tackle this problem?
thanks in advance