Pandora podcast
This week on the Inside the net podcast is an interview with Tim Westergren of Pandora. Pandora is a content-based music recommender/music streamer. In the interview, Tim enthusiastically describes how the Pandora and the underlying music genome project works. According to TIm, at Pandora, they have about 40 musicians that spend their days labelling songs with about 400 different attributes. It takes about 20 to 30 minutes for their trained musician to rate a song, and they are currently adding about 7,000 songs a month. After six years, Pandora has amassed labels for hundreds of thousands of songs and can do a find job of recommending music based upon similarity. However, they do have difficulties with scale. They are only able to add a very small fraction of new music each year and have to pick and choose which songs get added to their database. They skip some genres completely. For instance, there is no Classical music in Pandora. One of the advantages of a content-based recommender like Pandora has over the more traditional collaborative filtering models used in systems like last.fm is that they are immune to the popularity bias that is found in the collaborative filtering systems. A content-based system is just as likely to recommend a garage band as it is to recommend the Beatles, since it is immune to popularity, whereas a collaborative filtering system that is based upon user listening patterns is much more likely to ignore the unknown bands. They don't recommend unknown bands because no one is listening to them. Content-based recommenders can push listeners into the long tail of music, finding unknowns that sound like music we already like. Unfortunately for Pandora, the scaling problem makes them less likely to be able push people into the long tail. Pandora has to pick and chose which songs to process, and they will start with the most popular songs first, so they really can't push people too far into the long tail (yet). Despite these issues, Pandora is really good way to explore and discover new music. It is worth trying, and since it relies on flash, it runs on just about any platform out there..
http://tapestry.allmusic.com/
Posted by Jim on January 14, 2006 at 05:59 PM EST #