[ontIgnoreRealit]

1) Not every person wants the same kind of fit(snug,slim,skinny,relaxed,classic,natural, etc) or drape.

2) Look at this image to see the difference between different brands and their size 8 measurements.

http://www.nytimes.com/imagepages/2011/04/24/business/201104... via http://thesocietypages.org/graphicsociology/tag/womens-size-...

There's many variables to sizing. It wouldn't be impossible to create a database of measurements for every single piece of clothing.

And then there's the issue of shrinkage.

[cscheid]

1) Not every person wants the same kind of fit(snug,slim,skinny,relaxed,classic,natural, etc) or drape.

Those are irrelevant details. The basic form of input is "person A gives rating R to item B". This is precisely the formulation behind successful solutions to the netflix problem.

[creamyhorror]

You would need a pretty well-populated training set (assuming SVD or similar ML algorithm), I imagine, and somehow I think that might be difficult with so many unique item Bs. With Netflix you have many people watching the same movie, but I think there are many more dresses around, and women probably won't be rating hundreds/thousands of them (which is easier to do on movies).

[cscheid]

The Yahoo Music dataset has these characteristics (more items than users), and a combination of methods (including the SGD-based matrix factorization that people call "SVD" in recommender lore) did pretty well on it (KDD Cup 2011: http://kddcup.yahoo.com/).

[creamyhorror]

Interesting, thanks for the link.

What I find a little disappointing is that the top 3 prizes for this Netflix-Prize-style competition total $7,000 in value. Is technical brainpower, even that of students, really worth that little? Maybe startups should band together to hold their own optimization contests if the 'market rates' are so low.