In a study presented at the European Conference on Information Retrieval 2025, we show that uncertainty in emotion-annotated data cannot successfully be estimated with currently common machine learning approaches. While the estimation of model uncertainty is a relatively widely investigated area, less attention has been paid to estimating the distribution of highly subjective and thus potentially widely distributed annotations such as in ratings of valence and arousal for music emotion recognition. It turns out we really need better solutions!
Shoutout to my students Karn Watcharasupat, Yiwei Ding, Aleksandra (Teng) Ma, Pavan Seshadri. The paper is available online.
The Music Informatics Group is in the [Georgia Tech School of Music](https://music.gatech.edu], Georgia Tech College of Design.