Beating Spotify's algorithm: towards an improved emotion label for Billboard songs

Abstract

Spotify's recommendation algorithm tailors music offerings to create a personal listening experience. Though this recommender system performs admirably, there is always room for improvement. It remains unclear if Spotify accurately classifies the affective decomposition triggered by songs. This report tries to improve these emotion labels for Billboard songs. Emotion labels can be determined based on valence and arousal. The Spotify dispositions will be compared to the valence and arousal values derived from audio features, lyrics, audio features and lyrics combined, and a listener panel. These comparisons will provide insights about how emotion labels behave when audio features and lyrics are decomposed or combined. A survey was conducted to validate the results of Spotify. Participants had to rate the most "extreme" songs on valence and arousal inter alia. Results showed that it is necessary to analyse and combine valence and arousal values from the audio signal and lyrics. Based on the valence and arousal values of the mentioned models, significant differences were found compared to valence and arousal values provided by Spotify's algorithm. From the created models, it can be concluded that Spotify applied normalisation to increase the difference between emotion labels. This way, Spotify can provide a better recommendation based on emotion labelling. Compared to a combination of audio signal and lyrics values, Spotify did a fairly accurate job in labelling emotions.