I have been working on multiclass classification method.
I have an idea on how to extend this method such that it can be semi-supervised.
What are the best practices in evaluating semi-supervised methods?
I have two main questions:
- What data sets do people typically evaluate on?
- What metrics do people use for comparison?
I can imagine that people would just use typical data sets, such as iris or mnist and then use part of the data set as unlabeled.
Do people typically try out different proportions of unlabeled data?
Do they do this by keeping the number of labeled data fixed and adding more? Or do they vary the number of labeled data?
Is there any "accepted" standard way to do these evaluations?