I’ve been going through some of the FastAI course materials, and would love to implement a collaborative learner recommender in some work I’m doing, but I’m trying to figure out a meaningful way to put it into production and wanted some thoughts…
Unlike the tutorial example, my training dataset is quite small, and I expect it to develop over time - eg, I might have 100 entries now, but might have 1000 at the end of day 1.
The key question then is is there any good practice guidance on how, when and how often to retrain? I guess I could set the model to retrain every X hours based on new data, but I guess ideally I’d have some sort of reinforcement learning that tweaks the weights with every new input? Or is that likely to be a huge amount of training time?
Would welcome thoughts from anybody who has done this or similar!