For a time I had no hope that recommender systems like Amazon.com's "Recommended for You" section would be useful to me specifically. The predictions were often predictable. Buy a CD from artist A and get a list of the most popular CD's from that artist. Not useful.
Some time ago I came across Pandora.com, which is an adapting radio station, which chooses songs to play based on what songs you have added to a station and what songs you rate positively. I actually learned of several songs and artists I was unfamiliar with that I now like (such as "Question Everything" by 8Stops7). However, it does not play all songs that are similar to the songs I tell it. And some days I find myself disagreeing with all songs played.
I think that as time goes on recommender systems will improve and we will give some credibility to recommenders. Perhaps the Netflix prize will help in that regard.
Wednesday, October 29, 2008
Netflix Recommender System
Netflix is trying to motivate research in the area of recommender systems and on Oct. 2, 2006 offered $1 million to anyone that could improve upon their current recommender system by a specific measure (improve RMSE by 10%). Recently I took a look at the current standings and one team is very close (improvement around 9%). Interestingly enough they had a few papers showing how they do it.
Specifically what we are talking about is collaborative filtering. There are two main approaches, either you look for global patterns in the matrix of ratings or you use the ratings from similar items or users. BellKor (team name) was able to successfully merge these two ideas into a single solution that outperformed (at the time of submission) any other approaches using one of the two approaches.
What impressed me most about the paper I read (Factorization Meets the Neighborhood: a Multifaceted Collaborative Filtering Model) was that in addition to testing RMSE for the test, they tried to look at the users perspective. We want to know what movie to watch now. They compared other approaches against theirs on whether they would recommend in the top 5 or top 20 a movie you would watch and rate a 5. Well done. We should all keep the end user in mind.
Any one have a really good or bad experience with recommendations made by computers?
Specifically what we are talking about is collaborative filtering. There are two main approaches, either you look for global patterns in the matrix of ratings or you use the ratings from similar items or users. BellKor (team name) was able to successfully merge these two ideas into a single solution that outperformed (at the time of submission) any other approaches using one of the two approaches.
What impressed me most about the paper I read (Factorization Meets the Neighborhood: a Multifaceted Collaborative Filtering Model) was that in addition to testing RMSE for the test, they tried to look at the users perspective. We want to know what movie to watch now. They compared other approaches against theirs on whether they would recommend in the top 5 or top 20 a movie you would watch and rate a 5. Well done. We should all keep the end user in mind.
Any one have a really good or bad experience with recommendations made by computers?
Subscribe to:
Posts (Atom)