Ten Reasons Why

A recommendation network is basically the "People who liked/rented/bought this also liked/rented/bought _______" functionality you are seeing more and more in e-commerce sites. In essence, every social software network is a recommendation network. With social network sites like Friendster, Orkut (which I haven't been invited into! Wah!), Ryze, etc., there's typically an implicit recommendation in the establishment of "friendship" with another user.

Recommendation systems seem to fit a semantic web model quite well. The basic structure is

user ---> hates/dislikes/likes/loves ---> object

With social software networks it just so happens that the object is another user ("Joe likes Mary" or "Joe befriends Mary"). Other recommendation systems typically have users rating books, music, films, etc.

I like the recommendations I get from Netflix; they're nearly always spot on with my tastes. Amazon's recommendations are not as good and probably gamed by Amazon to drive purchases. E.g., if you say you've really liked a book by a particular author, immediately all of that author's other books float to the top of your recommendations. I suppose that's generally true, but it's seems like a cheap correlation, probably because Amazon recommendations are driven by purchasing as well as by ratings.

However, the great thing about Amazon's recommendations is that they are cross-product: it recommends the Lost in Translation DVD partially because I liked the short story collection, Cathedral, by Raymond Carver. Now Cathedral and Lost in Translation aren't even works in the same media, but there is definitely something Carveresque about Copolla's film -- that's a good correlation.

What I want is a social/recommendation network that has no boundaries. I want to know that people who liked Cathedral and Lost in Translation also dine at Ten Penh and buy groceries at the Whole Foods on P Street and are going to The Shins show at the Black Cat and wear Kevin Cole shoes and finds the girl named Jane who lives around the corner from me attractive. Sort of a crazy combination of uber-recommendations and social networking.

Stephen's proposed approach -- a marriage of blogging, FOAF, and some kind of metadata (e.g. dislike, like, love) -- comes closer to what I'm looking for. The challenge, I believe, is how to represent the object being rated, reviewed, recommended, or commented upon. I think his system's effectiveness may fall apart there because there's no common frame of reference, which is what non-distributed social network software provides.

By limiting the semantic frame of reference narrowly ("User A recommends User B" where all Users are uniquely identifiable or "User A recommends Book X" where all Users and Books are uniquely identifiable) you avoid fuzziness. In today's social networks, John can recommend Mary or John can recommend Cathedral, but John (for the most part) can't recommend a particular diner in Manhattan because it is neither a User or a Book -- it's outside the bounded frame of reference.

With Stephen's model I believe John could recommend another user, a book, or anything under the sun, but I suspect that if John has blogged that he likes Lost in Translation and pointed to a link to the DVD on Netflix and Mary has blogged that she likes the same film but linked to the database entry for it in IMDB, then Stephen's hypothetical semantic social network application may not be able to distinguish that John and Mary are in fact talking about the very same recommended object.

I suppose combing a Google-ish level of indexing and searching might get around that to some extent. Stephen suggest this: "It will be a search / aggregation tool that uses FOAF and RSS aggregators to satisfy queries based not only on the content of an article but on information about the authors of those articles." E.g. I think this means if John and Mary both use the text "Lost in Translation" in their semantically, socially networked blog entry, then maybe a search query can be used to semantically identify that they are both recommending (at least) the chunk of text, "Lost in Translation." However, there still needs to be a way to relate the text "Lost in Translation" with the film Lost in Translation (not to mention the "information about the authors"). For socially networked blog entries, this might work, but I don't think it makes the next step to the sort of "celestial recommendation network" (a la the idea of a celestial jukebox) that I'm dreaming about.

The challenge to a "celestial recommendation network" is one of (as seemingly always) metadata. Everything recommended needs to be identified. In Amazon they can do cross-vertical recommendations because Amazon has unique metadata (the ASIN number) for everything in all its web-based stores. However, outside a bounded, centralized system like Amazon, though, you lose the ability to uniquely refer to the object being recommended unless everybody previously agrees on a schema by which you will bind references to that object. And I think it's some approach to that kind of schema that is required to make a celestial recommendation network a reality.