3 ms·
He's not correct about the Netflix Prize algorithms. The leading ones all assume time-effects so that your rating on one day is not the same as on another day.
by larryfreeman 17y ago
He's not correct about the Netflix Prize algorithms. The leading ones all assume time-effects so that your rating on one day is not the same as on another day.
For the explanation of the winning team's view on time-effects as of 2008, see here:
http://research.att.com/~volinsky/netflix/Bellkor2008.pdf http://research.att.com/~volinsky/netflix/Bellkor2008.pdf
For a simpler example of a time-based global effect, (see User x Time Effect, for example):
http://algorithmsanalyzed.blogspot.com/2008/05/bellkor-algorithm-global-effects.html http://algorithmsanalyzed.blogspot.com/2008/05/bellkor-algor...
- durana 17y agoSome of the new models that came out of the Netflix Prize do take into account temporal changes. From a user perspective, it seems to be how a user's rating style can change over time and how a user's taste in movies can change with time. So the author is wrong when he says the models don't take into account temporal changes. The author suggests that the decision making process around what movie to watch in the future is different from the process of deciding what movie to watch right now. None of the models I've seen involved with the Netflix Prize try to take this into account as far as I can tell. I'm not sure how being able to make recommendations tailored to these two different decision making processes would be of value, but it would be different from the existing published models.
- ovi256 17y agoI would also recommend Yehuda Koren's (Bellkor leader, if I recall correctly) KDD 09 excellent paper, "Collaborative Filtering with Temporal Dynamics" http://research.yahoo.com/pub/2824 http://research.yahoo.com/pub/2824 which speaks exclusively of how to handle temporal dynamics.