Hi karndeep,

1. We try to represent user as sequence of movie genere he has interacted with. So we use the info to represent user and try to find his highly interest genre via cosine similarity score with movies. After all the user is nothing but the movies/genre s/he interacts with.

2. we can train a user tower model which incorporates all the user actions and convert it into vector. You can look for two tower models and how they are trained for more info.

3. Its subjective. We use faiss because of its optimized storage , fast retrieval and GPU capabilities. Anyone can setup faiss. but with ES there are some cost and sharding of data needs to be taken care. So i'll suggest list down the pros and cons from your use-case perspective and decide.