Collaborative Data Collection for Music Mood Classification

• Experiment with different acoustic features which may be more informative in discriminating valence or arousal. • Alter song clip lengths to determine the length that produces the most accurate classification results. • Investigate modifications to MoodSwings to make gameplay more entertaining in order to collect more data. With the growing size of personal digital music collections, the field of music information retrieval seeks to provide new ways of organizing, storing and maintaining these libraries. In developing automated tools to organize music by its emotional content or mood, the collection of “ground truth” data necessarily requires human subjects. We have developed MoodSwings, a game which records dynamic (per-second) labels of players ’ mood ratings of music