Matinee video & recommendations, open late
Rate films you've seen and the staff picks update as you go. Behind the counter is a model trained on – ratings from – MovieLens members. Your browser places you among them and ranks – titles. Nothing leaves this page.
a movie recommender, written from scratch
A one-star tape still tells the model what you watch. Stars only set how much it counts.
Each of 610 MovieLens members had a random 20% of their ratings hidden. Every model then picked 10 films per member from about 9,700 they hadn't rated, and a pick counted when the member had given the hidden film 4 stars or more.
The rating predictors were the wrong tool. Models built to guess star ratings had the lowest error ( RMSE), but their top 10s almost never held a film the member went on to rate highly. Just recommending the most-rated films beat them.
Learning from what people watch fixed it. Implicit ALS treats every rating as "watched this", with stars as confidence. Its picks land more often than most-popular's, and come from × more of the catalogue.
Checked against Spark. The Scala ALS is written from scratch and lands within of Spark MLlib's on NDCG@10. The batch job that writes every member's picks runs on Spark itself.
What isn't on the shelf. Only titles at least people rated, so every recommendation rests on enough data.
| MODEL | P@10 | NDCG | RMSE |
|---|
Ratings from the MovieLens latest-small dataset, F. Maxwell Harper and Joseph A. Konstan, "The MovieLens Datasets: History and Context", ACM TiiS 5(4), 2015. The film factors on this page are derived from it and shared under the same GroupLens terms, for non-commercial use. Titles only; no cover art.