The validation data provides an estimate for the threshold. We can define a accuracy measure or a condition, for example, maximum f1-score. And find the threshold that gives the maximum f1-score. Note that we should do this on the validation set and not the test to ensure we do not overfit.