Hacker News Reader

Comment by q-big | Hacker News Reader

> Well, you say that, but isn't that exactly why Netflix got rid of ratings?

> All the documentaries were getting really high ratings, so would display highly in searches, but not many people actually watched them.

This rather shows that ratings do work, but are used wrongly for giving recommendations:

If you want to create suggestions for a user, in many case the wrong answer is "suggest what has really high ratings", but rather "given the ratings that this user gave and the films that he watched, what will he also like."

The fact that these documentaries get high ratings (the same might hold for art house films) shows that there is some (niche?) audience which really loves this kind of films, but not that "John Doe" will love it, too.

View on HN

Predicting exactly this was the premise of the famous Netflix Prize, one of the first open machine learning challenges https://en.wikipedia.org/wiki/Netflix_Prize

I remember my professor talking about the ensemble approach the competitors were taking in a data mining class I took. I was saddened when Netflix ended up not using it because it was too slow and expensive or something.

It didn't promote Netflix content above third party content. Not enough control for the algorithm is unacceptable for the marketing guys.

Thanks! I'd totally forgotten the real reason.

I remember when that competition was released, I was working (as an undergrad) in GroupLens, which had MovieLens still around but it was getting a bit of bitrot.

It was always weird to me that GroupLens didn't spin up a team for it, but it seemed like everyone in the research group had moved on to other things and didn't want to context switch back. Someone mentioned something like "shame they didn't do this 10 years ago, a million dollars would have been nice". I think someone was doing tagging on movielens, but I don't remember the details.

I got the sense that neither Riedl nor Konstan (or any of the current grad students it seemed) wanted to pursue it (Terveen, I think, wasn't in to recommender systems at all in the first place).

I don't think the lab had any funding problems haha, so maybe that was what it came down to.

> This rather shows that ratings do work, but are used wrongly for giving recommendations

And I know I'm (possibly) a minority on hacker news, but I prefer the new system. I was giving everything I wanted to watch more 4 or 5, even when it was clearly not the case, but because I want recommendations of things I'm going to like AND actually watch

Isn’t this why it now says something like “your match: n%” now?

Reply on news.ycombinator.com