xlvector – Recommender System

After Netflix Prize, Netflix inc. announce that they will hold another competition by using more user data (gender, zipcode, age, etc) to improve prediction accuracy of recommender system. However, many Netflix users think this is violate to they privacy rules and do not want Netflix to reveal such data.

http://www.wired.com/threatlevel/2009/12/netflix-privacy-lawsuit/

As a competitor of Netflix Prize, I also agree with these users and I think we do not need more user data to improve accuracy. I think, we should focus on different problem. Netflix Prize use RMSE to measure the quality of different recommending algorithms. However, in real systems, our main task is recommendation not prediction.

In real recommender system, we have to recommend a list of product a user may like. This task is different from the problem solved by Netflix Prize. The problem of Netflix prize is to predict the rating a user will assign to a movie. In this competition, we already know what movie a user is watching and our task is only to predict rating. In this way, Netflix Prize does not solve the main problem of recommendation: how to recommend a list of items to active user?

Further more, users do not only rating items in a system, they may view/buy/comment items and rating behavior is only a small part of user behaviors. So, I think, Netflix do not need to reveal user data, they can reveal more behavior data rather than rating.

Netflix can also reveal nothing to hold another competition only data revealed in NetflixPrize1. They can propose different problems, such as how to recommend a list of items for an active user. One disadvantage of Netflix Prize is that, after Netflix Prize, most of researchers in recommender system thinks rating prediction is the only thing in recommendation. I think, Netflix should not focus on rating prediction now. They should promote research in other recommendation problems, such as Top-N problem, diversity, novelty, etc.