Recommender systems

Last week there was a workshop in Bilbao on recommenders (Recommenders 06) and I've been reading the blog that tracked the sessions. Some of the sessions that looked particularly interesting were:

Building a personalised radio station
John McPherson on building a personalised radio station using the MyStrands API, it looks very similar to Pandora and last.fm personalised radio stations. MyStrands is looking particularly interesting because it provides what looks like a very comprehensive and open web API to its data, including access to their music catalogue, search, recommendations, playlists and tagging. Another session included a paper on how the MyStrands recommendations work - basically a content-agnostic recommender that pre-computes a network of associations and correlations between all the items in the system, with customisable metadata for different applications to allow the customisation of searches - e.g. only jazz, only stuff available at iTunes, only stuff in my library...

Playlist visualisation
MyStrands also showed a cute little visualisation of your playlists with associated recommendations. I'd have to question the visualisation of the bebop mix which features Dexter Gordon surrounded by the Beastie Boys, Squarepusher and Peter Gabriel and Giant Steps initiating the discovery of Cecil Taylor, Pharoah Sandres and Archie Shepp. I suspect they haven't really got much data on this area of music but it does look gorgeous.

Secure recommendation systems
Bamshad Mobasher’s talk on secure recommendation systems was about how these systems can be attacked or gamed to deliberately promote, or demote, a particular piece of content and also how to protect against such attacks.

David Jennings was at the workshop and has written some notes up at Net, Blogs and Rock'n'Roll. I completely agree with him that recommenders are just one input to helping us find things. They are not magic systems that bring us whatever information we need before we even know we need it. The Duke Listens blog from Paul Lamere at Sun Labs is another good place to keep up with this area.

From personal experience I'm reasonably impressed with Amazon's music recommendations, though I do buy a fair amount from them so they've got a good amount of data. Most things they recommend I have already heard of though I would say I don't yet trust it as a recommender service. But certain CDs seem to regularly come up in my recommendations and I have noticed that my interest will be piqued if I see that same release talked about elsewhere. I've also written previously about how Amazon recommendations seem to follow aggregations from particular reviewers or sources. I mentioned this before with respect to Later... but I've also noticed this phenomenen from the Guardian's jazz reviews - add a reviewed CD to your shopping basket and other CDs reviewed that week are suggested. It seems that the reviewer or tastemaker ends up acting as an implicit but strong input to the recommender via the customers. You could see this as negating the need for recommenders if all they are doing is passing on mainstream media reviews but I think that what you are seeing is recommenders acting as aggregators of the mainstream media, the tastemakers, the customer/audience behaviour and all the other, possibly hidden or indirect, inputs to the recommendation system. Which, I think, is how they should work.