We propose an approach for the automatic creation of mood playlists in the Thayer plane (TP). Music emotion recognition is tackled as a regression and classification problem, aiming to predict the arousal and valence (AV) values of each song in the TP, based on Yang's dataset. To this end, a high number of audio features are extracted using three frameworks: PsySound, MIR Toolbox and Marsyas. The extracted features and Yang's annotated AV values are used to train several Support Vector Regressors, each employing different feature sets. The best performance, in terms of R2statistics, was attained after feature selection, reaching 63% for arousal and 35.6% for valence. Based on the predicted location of each song in the TP, mood playlists can be created by specifying a point in the plane, from which the closest songs are retrieved. Using one seed song, the accuracy of the created playlists was 62.3% for 20-song playlists, 24.8% for 5-song playlists and 6.2% for the top song.

This work was supported by the MOODetector project (PTDC/EIA-EIA/102185/2008), financed by the Funda- ção para Ciência e Tecnologia - Portugal.