Tuesday, June 14, 2005


Gnod is fascinating: it attempts to use AI techniques to identify similarities between books, films and music. There are very few details about how it works, but it claims to be driven purely by user feedback. But based on the way the entities dance around the music map, my guess is that it's powered by a Kohonen network. If so, it's the first really useful example of automatic clustering and unsupervised learning techniques that I've seen.

Gnoosic, the music component of Gnod, can suggest bands that are similar to bands that you already know about. Even better, the Music Map component can generate a visual representation of how close various bands are to a source band. As a quick test, generating the map for U2 gives R.E.M, Police, Inxs and Sting as nearest neighbours, which seems pretty good.

Gnod can also generate maps for films (Gnovies) and authors (Gnooks). The data for authors seems pretty good (looking for Arthur C Clarke finds Azimov, Douglas Adams, Orson Scott Card and so on), but the one for films seems to be a bit random (searching for Toy Story returned Once Upon a Time In America as the closest match)


Blogger Jonathan Thorpe said...

Hi Steve. Found this blog by way of Rupert Rawnsley's blog. Gnod looks interesting, but I think it's using a much simpler technique than a Self Organising (Kohonen) Map. My guess is that it codes each band/movie/book according to user preferences to form a feature vector, and then simply finds the n nearest neighbours to each item using some distance metric (probably Euclidean).
The visualisation is essentially a 'spring' graph that visualises those distances. Each item in the graph is connected by a spring (allbeit invisible) with the tension of that string is determined by the distance. The graph is then iteratively optimised according to those tensions between items to attempt to preserve the distances witin the graph. That's why the graph moves the way it does (like a bunch of things connected by springs!).
If a Kohonen map was used the visualisation would be much more static, and possibly not as good.

12:34 AM  

Post a Comment

<< Home