Posted on 12 December 2012

Part 10 in a series of videos recorded from ACM MIRUM 2012 in Nara, Japan.

Classification of songs by their emotions is a difficult problem because emotion is a subjective characteristic -- different people may label the same song with different emotions. To customize an emotion classifier, you need a lot of training data for each individual user. Collecting enough annotations to build an accurate and customized classifier is probably too burdensome for most users.

Dan Su presents a method for emotion classification which alleviates the amount of annotations required by the user to customize the classifier. By using a model from machine learning known as active learning, the proposed method only requests of the user the most informative instances of training data, thereby minimizing user participation during training while maintaining the classifier's accuracy.