3 ms·

Ok, the author is a bit of a moron. He said stupid things and detracted from what I am guessing his point is. He also didn't read about the Netflix algorithms

Ok, the author is a bit of a moron. He said stupid things and detracted from what I am guessing his point is. He also didn't read about the Netflix algorithms and how they actually work (all around terrible job of reporting indeed.) However, to circle back, what I think is his point is interesting: how do you model "mood over time" and "non-static behavior". You say they "can be accounted for" and don't talk about how. In the winner's documentation they say they treated "user preference" variation over time the same as chunked up "user bias" over time (they define these terms in the paper.) While that obviously performed well numerically, isn't it believable that this is not an ideal solution, but one done under severe time constraints? The problem is by no means solved...that's all I'm saying.

roundsquare 17y ago
If we just take his examples though, then the problem is on its way to being solved. So he puts movies on his queue that he doesn't watch? Great, he should give them low rankings. If the algorithm is good, that problem is solved. He likes some movies at night, some on the weekends and some with his wife? Well, okay, rate the movies how you thought they were when you saw them. The algorithm will recommend some movies for each situation. Your preferences change over time? Thats a bit harder, but was also out of the scope of the challenge. They were given static data. If you want to account for changes over time, then give the ratings some weighting that decreases over time or only use the last x months of data. But, like you said, he probably doesn't know much about the algorithm.
brg 17y ago
I agree that the author detracted from what could be a more interesting (less noisy) conversation by attempting to be affected by the "problem." By "can be accounted for" I mean that while posing a problem, the considerations do not yet prove to be insurmountable. Especially when the author referred to any algorithm used by Netflix, not simply one constrained by the contest's static training set and evaluation procedure. But we are in agreement that it is by no means a solved problem, and that's what makes this a great period for machine learning.

roundsquare 17y ago

If we just take his examples though, then the problem is on its way to being solved. So he puts movies on his queue that he doesn't watch? Great, he should give them low rankings. If the algorithm is good, that problem is solved. He likes some movies at night, some on the weekends and some with his wife? Well, okay, rate the movies how you thought they were when you saw them. The algorithm will recommend some movies for each situation. Your preferences change over time? Thats a bit harder, but was also out of the scope of the challenge. They were given static data. If you want to account for changes over time, then give the ratings some weighting that decreases over time or only use the last x months of data. But, like you said, he probably doesn't know much about the algorithm.

brg 17y ago

I agree that the author detracted from what could be a more interesting (less noisy) conversation by attempting to be affected by the "problem." By "can be accounted for" I mean that while posing a problem, the considerations do not yet prove to be insurmountable. Especially when the author referred to any algorithm used by Netflix, not simply one constrained by the contest's static training set and evaluation procedure. But we are in agreement that it is by no means a solved problem, and that's what makes this a great period for machine learning.