"We think you'd also like..." and the Math of Suggestion – Part 1 (from Directed Edge)
gruenderszene.de
gruenderszene.de
http://www.igvita.com/2007/01/15/svd-recommendation-system-i...
The article is definitely targeted at being a quick introduction to how recommendations work rather than being a how-to. Most Gründerszene's subscribers aren't developers so it's an attempt to present the stuff in a digestible format.
There is no doubt recommendation engines are a great sales generator. But are they any good at what they do? Plotting points on a graph based on user ratings has always seemed artificial to me. I may like a book because of its rhythm and someone else may like the subject matter; will we like the same books? This is why I think netflix created their competition. They're not looking for better math, but for a better approach all together.
At the end of the day, it probably has to be a combination of factors. For example, think of pandora's engine. It is based on analysis of the music, not just user preferences. And yes, math will be involved, which is why part two of this article will be worth reading.
http://cacm.acm.org/blogs/blog-cacm/22925-what-is-a-good-rec...
It's a valid question to ask, "Will these algorithms really suss out my personal taste?" And I think the answer is, a lot of the time, they'll get close enough to be interesting.
When we started Directed Edge, one of the tipping point moments was seeing how bad some of the related films shown in IMDB for my favorite film were. It was a real watershed moment this week when I was running our algorithm against the Netflix dataset after adding 15 of my own ratings to the system and my two favorite films (which I had not rated) came out at #1 and #5, respectively, and at that, with 300x less computation time than a standard item-to-item recommender requires. (We've got implementations of a lot of standard algorithms that we use for benchmarking our own.)