CONCEPT-BASED EXPLANATIONS FOR MUSIC EMOTION RECOGNITION

Concept-based explanations have been shown to be useful in other domains and recently found their way into music information retrieval research. The idea is to provide the user with human-understandable, high-level concepts as an explanation for a model's prediction. There are two ways of obtaining concept-based explanations for an existing model (post-hoc): one requires additional training data labeled with the concepts of interest, the other extracts concepts without requiring prior knowledge, which is generally advantageous; however, it suffers from the problem that we do not know the meaning of the discovered concepts and require users to label them. In this work, we study this problem in the context of music emotion recognition. We conduct a listening experiment which shows that extracted concepts satisfy desired properties: coherence and meaningfulness; demonstrate how concepts can be presented to users, and explore the possibility of automatically labeling concepts without requiring users. In addition, we show quantitatively that the concepts are also faithful to the model.