A Novel Graded Multi-label Approach to Music Emotion Recognition

Abstract

Music Emotion Recognition (MER) has become one of the key researched axes in Music Information Retrieval. Its main objective is to automatically recognize the effective content of music pieces. In this paper, we are interested in the task of Music Emotion Classification which success has plateaued in recent years. We wish to offer a new perspective by approaching the problem as a graded multi-label learning problem and therefore bridging the existing limitations presented by the categorical taxonomy in Music Emotion Recognition. In order to assess the suitability of this setting, we adapted a state of the art MER dataset by annotating it according to a graded multi-label format. Our initial studies conclude the promising potential of this approach.