I am interested in using recommender systems model created by @jeremy, but I want to optimize it for predicting the average rating of a movie (say a year in future when it has had n more viewers, assume I have a separate model that predicts the n users who will likely watch each movie in the future). Currently, the model predicts the rating of a given movie by a given user. While this model does great at that level of granularity, when you aggregate the ratings to create a per movie rating, the model accuracy goes down (at least for my dataset which is different from movielens).
I am guessing changing the network topology so that we directly predict the movie rating will be more accurate. Any ideas on how to change the network structure for this?