I have a classification task in which classes exist within a directed graph. That is a class may have subclasses which share an is-a relationship with their parent class. Now, I have a relatively large number of labeled samples for the parent classes, but as I travel through the graph to subclasses (which are more specific) the labeled samples available become more and more sparse. I want to apply the concept of transfer learning to predict the subclasses. Particularly, I was thinking of training a classifier to predict the parent classes first. Then using the pretrained model, I want to perform fine-tuning using the same samples, but this time labeled with their respective subclasses instead of the parent classes. Is this a valid technique? I'm concerned that this would somehow leak data. Does anyone know of any literature in which this approach is discussed?