Increasing Complexity of Problems

by rrusczyk, Nov 23, 2008, 9:41 PM

A little while back, I posted about the increasing complexity of problems. Basically, I think it's not clear which is increasing faster: our ability to solve problems or the complexity of problems we face. This article, about the Netflix problem (which is not unrelated to some of the problems we face in building Alcumus), is a great example of how hard it is to make small progress, and yet how valuable even incremental progress can be. (I also think it's another great example of using prizes to marshal resources to a problem, like the X Prize. I wonder if such a prize could be structured for education...)

news

Comment

4 Comments

The post below has been deleted. Click to close.

This post has been deleted. Click here to see post.

The Netflix competition is actually quite a famous problem for collaborative filtering in the machine learning community. Many of the algorithms which performed best in the competition are variations of simple ones as kNN. As you say the question is whether the 10% of improvement is actually possible or whether there is too much intrinsic noise in the data. Just imagine allowing your kids to choose movies or taking your mood swings into consideration the preferred choice and subsequent evaluation of the movie may exhibit strong variability.

For your work on Alcumus you may also want to meet up and discuss with
Yoav Freund from UCSD who is an eminent researcher in the field of machine learning and got know for ideas as boosting.

by orl, Nov 24, 2008, 10:53 PM

Report

The post below has been deleted. Click to close.

This post has been deleted. Click here to see post.

Thanks Richard, that was a really neat article. I found it amusing that my favorite film, 'The Life Aquatic with Steve Zissou', is on that list (of difficult-to-recommend films) -- it seems like even Wes Anderson fans can't decide whether to like it.

by blahblahblah, Nov 27, 2008, 8:45 PM

Report

The post below has been deleted. Click to close.

This post has been deleted. Click here to see post.

I've watched every one of the hard-to-recommend films, and my reaction to them has ranged from "Awesome" to "Awful". I can really see how it would be very tough to model whether or not someone would like them, and yet I think modeling these films is the most valuable prediction a model could make. The genre films are easy -- everyone knows whether or not they'll like <fill in cliched film here>. It's these more unusual indie films that I have trouble knowing whether or not I'll like. And yet they're the films that I'd most prefer to watch, since I'm usually 95% sure I'll hate <fill in cliched film here>.

So, I think these tough films are the big problem, and it doesn't look like much progress is being made (at least judging from Netflix's accuracy for me!)

by rrusczyk, Nov 28, 2008, 4:20 PM

Report

The post below has been deleted. Click to close.

This post has been deleted. Click here to see post.

It's probably inevitable that there are limits to how well one can predict films of that type. I wonder if using social networking tools like facebook or myspace to obtain a large volume of data regarding people's favorite movies would be a step in the right direction here.

by blahblahblah, Dec 1, 2008, 10:08 AM