Hello, I watched the video 5 from fast.ai course Part 1 about collaborative filtering.
I have understood the concept, but when I want to use the model in the real situation, when there is a new user, I have to recommend the movies according to the ratings he/she made. Similar for new movie.
Many deep learning models are feed-forward, so when I give it a new data, it pops out a result(prediction, classification, etc). But the model from lecture 5 first need the embedding vector for a new user or new movie first. Is there a way the collaborative filtering handles this new data issue, or should I use another model?
I think this is a good question. I’d like to know the answer, too.
This is called the cold start problem. In general, you would have another model that combines with your collab filtering model or “jumps in” when collab filtering cannot help.
Are there any resources related to DL applications?
Deep learning for recommenders is a huge and complex topic. The Youtube Recsys channel from 2016 is a good place to start. Paul Covington’s recommender from youtube was one of the first.
There are a number of ways to solve cold start, either through projecting the new user into the latent space based on early actions, starting them from a popularity based model, or by representing all users in item space as some aggregation of the latent representations of items they view.
It’s a huge research area though, and not a lot of intro material is available.