Netflix Prize: Forum

This contest has produced some methods that give a better RMSE score to the predicted ratings of users of Netflix.  I would like to submit that although that is a good thing, there are other things they could do that would produce better recommendations for their users.

My first observation (made here before) is that the contest did not use a lot of data that users had supplied in their previous ratings -- the category of "not interested".

My second observation is that it may well be that the RMSE score improvements are coming from the middle of the users ratings -- mean, median or mode, take your pick.  However, what would be of most benefit to Netflix is for them to be able to make reliable recommendations about movies that people will rate as 4s or 5s.   Most such recommendations should not care about predictions of movies that the user would rate as a 1 or 2 or even as a 3.

For example, instead of using RMSE  which is based upon (score - prediction)**2, how about using a score based upon weight(score) * (score -prediction)**2, where weight might be linear, quadratic or even exponential.  Such scores would give more strength to the highly rated movies than is being currently done.

A related third observation is that when Netflix shows a user the "recommended movies", it show him the top movies, even if the predicted score is less than 4.0.   Because I am fairly frugal with handing out 4 or 5 ratings, Netflix never recommends any movie to me in the general category.