There's a session-by-session blog of the whole event, with downloadable presentations from many of the speakers. Sadly I don't have time to record detailed observations on what I learnt there, though several points will all feed into what I write from now on. Certainly I know a lot more about the different varieties of systems for recommending music, films, digital cameras, Mercedes (statistically. people who bought leather gloves also bought Mercedes, apparently), as well as their strengths and weaknesses. I recommend Juntae Kim's presentation as an introduction.
Most interesting to me was John Riedl's talk and subsequent discussion about the impact of recommender systems on community. I'm wary of a view of recommender systems that believes they are complete solution. Recommender systems are a form of push technology, and as such they could be seen as a systems designed for sheep — when I argue that we need systems designed for foxes, squirrels and seagulls (in other words, active foragers not herd-followers). Any idea that recommender systems will do all our filtering for us, and all we need is a highly-personalised relationship with some well-crunched data needs to be nipped in the bud.
At the end of the day, the recommendations that come out of any system are just one 'input' among many for music/film/camera fans.
Riedl talked most about the community context and application of recommender systems. He made the strong claim that the Social Web (a.k.a. Web 2.0) will be driven by recommender systems. What's interesting is that he seemed to be talking most about a kind of second-order recommender system: not a system that recommends what music/films/camera you should buy, but a system that recommends what kind of recommendations you are best placed to make to other people. The issue for Web 2.0 resources like Amazon reviews, Wikipedia entries, and Rate Your Music (not to mention MyStrands, Yahoo Music, Last.FM etc) is how to increase the proportion of active participants in their communities.
Using a collective effort model from Karau and Williams (1993), Riedl suggested that many people won't edit a review entry for a film like Titanic because it's so popular that they assume someone else will. They're much more likely to spend time on niche films that they've seen but others may not have done. So Riedl proposed a kind of 'SuggestBot' to recommend items which Wikipedia/other tasks you should work on. This would mean that the allocation of work in creating a kind of review-and-recommendation commons could be optimised. Then this resource would be very valuable for the foxes, squirrels and seagulls.
I discussed this more with John Riedl in one of the breaks, and he pointed out that there has been a research question hanging around for a while (he dated it back to the CSCW '94 conference): when do you create patches for people to forage for discoveries, and when do you let them grow their own patches.
He also hailed recommender systems as a means to 'democratise' the process of discovery — to take it out of the hands of the mainstream media. But, as was pointed out elsewhere, we need to be careful about what 'democratise' means here. We don't want a tyranny of the majority. The Rhapsody charts apparently have a 'Jack Johnson effect', where so many of their users like his music that, almost regardless of how you specify your preferences, you will get recommended Jack Johnson sooner or later. Alvy (I think) notes how MyStrands had a similar effect with Coldplay recommendations recurring too frequently, while Paul Lamere notes how in Last.FM it was the more indie Postal Service who dominated the charts (and thus many recommendations) for an age. Chris Anderson went as far as suggesting a "forced decay of bestseller lists" to counteract this effect. (Better not call them 'bestseller lists' any more, then.)
The picture at the top is of Claudio Baccigalupo presenting his research on a tool to recommend the order of playlists containing a particular song, using an Artificial Intelligence technique called case-based reasoning. All those playlist-sharing systems on his slide are ones that I have reviewed previously. You can try out this tool at MyStrands Labs. I put in I Don't Like Mondays, and, sure enough, it came up in a playlist preceded by Loving You Sunday Morning, Sunday Morning, and Seven Sundays. I guess Claudio's case base extends beyond just MyStrands playlists, since no one seems to have played I Don't Like Mondays there yet.