I'm using self-training as a semi-supervised approach to increase the size of the set of labeled observations. In each iteration a classifier is built based on already labeled observations and then those examples that have been classified with highest confidence are added to the labeled set. Then in the next steps a new classifier is constructed based on the new set and this procedure is iteratively continued.
I would like to prove that increasing the number of labeled observations improves the accuracy of the classifier in each iteration. What are the possible ways to evaluate this given that my initial set of labels is quite small. I could extract the validation set but I think this would lead to having too few examples used to expand the labeled set.