An interplay between genre and emotion prediction in music: a study in the Emotify dataset

Abstract

Automatic classification problems are common in the music information retrieval domain. Among those we can find the automatic identification of music genre and music mood as frequently approached problems. The labels related to genre and mood are both generated by humans, according to subjective experiences related to each individual’s growth and development, that is, each person attributes different meanings to genre and mood labels. However, because both genre and mood arise from a similar process related to the social surroundings of an individual, we hypothesize that they are somehow related. In this study, we present experiments performed in the Emotify dataset, which comprises audio data and genre and mood-related tags for several pieces. We show that we can predict genre from audio data with a high accuracy; however, we consistently obtained low accuracy to predict mood tags. Additionally, we tried to use mood tags to predict genre, and also obtained a low accuracy. An analysis of the feature space reveals that our features are more related to genre than to mood, which explains the results from a linear algebra viewpoint. However, we still cannot find a music-related explanation to this difference.