Imagine you want to predict 2048 classes. Instead of asking one model to predict all of them at once, is it a known type of solution to have a model predict which cluster or group of classes an input belongs to first, and then have a secondary model that predicts the class from the cluster?
This would end up with n+1 models, where n is number of clusters. I imagine you could create clusters manually by looking at a baseline model’s confusion matrix, or by unsupervised learning results like k-means.
I tried searching for this type of model but couldn’t find anything. Does this exist? Is it a terrible idea? Is it good for some applications only? Please let me know.