The point of having groups of content is to combine traditionally dissimilar types of content-- movies, music, housewares, etc...
For example I can make a group of content called "new living room" and add all of the things that go into my new living room. This includes the music and movies that I have stored there, the type of TV I bought, the type of couch, stereo receiver, speakers, or even the paint on the wall. When someone searches for something within that collection, the system knows that someone else, somewhere, has combined that "thing" with the other "things" in the collection, so they get rated higher as being compatible.
This page explains many of netflix's limitations well: http://harry.hchen1.com/2006/10/03/391. But more importantly, look at these limitations in light of how the discovery engine is organizing its preference data and how it's collecting preference data.
The critical difference with the discovery engine is the idea of a group of content that users fill themselves with content based on criteria they see as relevant. Yes, a users aggregate preference composition is important, but what is more important is their set of preferences regarding a specific collection of content. This way, a user can be really into classical music, horror movies, and modern furniture, and get relevant recommendations for each interest, connecting with people who are most in the know regarding each interest.
I'm not sure what you're asking here-- Netflix doesn't care about anything but movies, and it probably wouldn't be able to recommend movies any better if it knew your musical tastes, or even how your tastes compare to mine.
The idea is that if you search for "sony SSK70ED," on the "discovery engine," it will show you what other people have paired with those speakers, such as receivers, furniture, and televisions. In a way those things are "similar" to the speakers because they complement them. Of course, the system shows you similar speakers first, but the complementing items are interesting results to have when you're searching for a specific item.
> I'm not sure what you're asking here-- Netflix doesn't care about anything but movies, and it probably wouldn't be able to recommend movies any better if it knew your musical tastes, or even how your tastes compare to mine.
That's wrong. The only thing that is movie-specific about the netflix recommendation system is the preference data that it runs over. It doesn't know movies from eyebrows.
If netflix (the company) also collected preference information about music, the recommendation engine would predict music preferences. And, since it would have both music and movie info, it would use music prefs to recommend movies and the reverse, just as it uses movie prefs to predict movie prefs today.
Amazon's "users who bought {something} also bought" is an example of "doesn't know anything about the domain". (They have to tone it down to keep it from recommending "strange" things that are way out of category.)
Disclosure: I know the guy who implemented NetFlix first recommendation system and have written a collaborative filter myself. I know what I can do with the fact that we both like the Pogues and Chunky Monkey. I still don't see what I can do with how we group those preferences.
"I still don't see what I can do with how we group those preferences. "
The way in which this site will allow users to group their preferences seems like a slight organizational difference when compared with other recommendation sites that use collaborative filters, but it has huge implications. This post is meant to give people a taste of what I'm starting to try and find others who are interested. I'd love to talk about any and all specifics and their implications especially if you are programmer. If you want to chat my AIM is rocksld3.
For example I can make a group of content called "new living room" and add all of the things that go into my new living room. This includes the music and movies that I have stored there, the type of TV I bought, the type of couch, stereo receiver, speakers, or even the paint on the wall. When someone searches for something within that collection, the system knows that someone else, somewhere, has combined that "thing" with the other "things" in the collection, so they get rated higher as being compatible.