Recursively Train Ensemble Methods

[mo-fu]

I think it would be nice if calling annif train on an ensemble method, the sub classifiers would be trained as well.

[osma]

Thanks for the idea.

I think this should be an optional feature. In my view, ensembles are best used to correct biases in basic (non-ensemble) methods. The basic models can be trained with large amounts of (possibly biased but readily available) short text metadata; then the ensemble can be trained with a small amount of high quality full text documents.

But of course there are different usage scenarios and different kinds and amounts of training data.

In any case, when using the learn method on ensembles, they should probably also pass the input to underlying backends (if they support learning - most don't currently).