I need to perform classfication of hundreds of classes. New classes arrive regularly. I also have some large training set (thousands of samples).
sklearn trains all its underlying classifiers with the same hyperparameters. My idea is that I can train one classifier for each class separately, when class arrives. This should give the following benefits:
- No need to refit all previous classifiers. Given training data is large enough, one additional class doesn't make big difference for what's already trained. Training time doesn't grow linearly on classes.
- Hyperparameters are chosen separately for each classifier which should improve accuracy.