Talk:Audio Tag Classification

Hi folks, Concerning the discussion on whether each tag should be treated as a separate binary classification task, or not, I propose presenting all tags associated with a track at once, since this may carry more information than separate binary classifications. Having each training example belong to more than one from a set of labels is often called multilabel classification in machine learning. It has been shown that better results can be achieved if one uses algorithms that take into account the corellation of labels, instead of treating each label independetly. We have a related paper on this subject in ISMIR 2008. See our multilabel classification page for more information (http://mlkd.csd.auth.gr/multilabel.html). It would be interesting to study whether the above claim will hold in this tag dataset as well, or if independent classifiers will perform better. Best, Grigoris