ABSTRACT
One of the key challenges in music information retrieval is the need to quickly and accurately index the ever growing collection of music on the Web. There has been an influx of recent research on machine learning methods for automatically classifying music by semantic tags, including Support Vector Machines [12, 13], Gaussian Mixture Models [18, 19], Boosting [2], Logistic Regression [1], and other probabilistic models [6]. The majority of these methods are supervised learning methods, requiring a large amount of labeled music as training data, which has traditionally been difficult and costly to obtain. Today, the shortage of labeled music data is no longer a problem. There is now a proliferation of online music Web sites that millions of users visit daily, generating an unprecedented amount of useful information about each piece of music. For example, Last.fm, a collaborative tagging Web site, collect on
the order of 2 million tags per month [8]. Without prompting, human users are performing meaningful computation each day, mapping music to tags.
