Hybrid Recommender: Adding new users and items to an existing model

[sgaseretto]

Does LightFM support a way to initialize and add embeddings for new users and items to an existing (pre-trained) model, so that we don't lose all the pretrained embeddings and also being able to index those new users and items, without having to build and train a completely new model from scratch.

[mwojnars]

...Or, to reframe the question, does LightFM support (or will support) incremental training, that is, training a model on an (unbounded) stream of events, such that at every step, the model can be used for predictions, even before the training completes (possibly never).