Content-based recommendation with the candidate tower in retrieval model

Hi, I’m studying the two-tower model and I was wondering is it possible to do the content-based recommendation with the candidate tower in a trained two tower retrieval model.
For example,
I’ve trained a two tower model which have a user tower and a movie tower. Then I can get the movie embeddings from the movie tower. Then I just calculate the similarities of the embeddings of each movies ( or use sth like scann) to find the top-k most related candidate movies of each movie.
Is it correct and reasonable?

Hi @Wenhui_Jolie_Zhang ,

This approach is correct and reasonable because it uses the information that the movie tower has learned about movies to recommend movies to users. The movie tower has learned to represent movies as embeddings that capture the content of the movies. By calculating the similarities of the embeddings, you can find movies that are similar in content to the movies that the user has already interacted with.

This approach is also efficient because it only requires you to calculate the similarities between the embeddings of the movies in your dataset. You do not need to retrain the model to make content-based recommendations.

I hope this helps!