Electrical engineers at the University of California, San Diego have recently been pitting Apple’s Genius music recommender program against their self-proclaimed, new and improved version, which is said to include an ignored sector of music, dubbed the ‘long tail’, in music recommendations.
It’s always been well known that radio suffers from a popularity bias, in which the most popular songs receive an inordinate amount of airtime, while less popular, sometimes better music is heard very rarely. In Apple’s music recommender system, iTunes’ Genius, this bias is magnified. Genius uses “collaborative filtering” on purchase statistics from iTunes Store- they’ve sold over 6 billion tracks- in order to help people organize their music and discover songs and artists they’ve never heard based on similarity to a “seed” song that they have liked in the past. But an underground artist will never be recommended in a playlist due to insufficient data from the store. It’s an artifact of the outdated popular collaborative filtering recommender algorithm, which Genius is based on.
In order to establish a more holistic model of the music world, Luke Barrington and researchers at the Computer Audition Laboratory have created a machine learning system which classifies songs in an automated, Pandora-like fashion. Rather than having humans explicitly categorize individual songs, this sysyem captures the wisdom of the masses using a Facebook game called Herd It, and then uses the data to teach statistical models.
This machine can ‘listen to,’ describe and then recommend any song, regardless of its popularity or lack thereof. The more people that play the game, the more data is processed and the better informed the machines get. and based on user feedback, their experiment with automatic recommendations produces playlists that are at least as exceptional as Genius, especially for recommending undiscovered music. Courtesy of slahdot.org