A Study on Cross-cultural and Cross-dataset Generalizability of Music Mood Regression Models

The goal of music mood regression is to represent the emotional expression of music pieces as numerical values in a low-dimensional mood space and automatically predict those values for unseen music pieces. Existing studies on this topic usually train and test regression models using music datasets sampled from the same culture source, annotated by people with the same cultural background, or otherwise constructed by the same method. In this study, we explore whether and to what extent regression models trained with samples in one dataset can be applied to predicting valence and arousal values of samples in another dataset. Specifically, three datasets that differ in factors such as cultural backgrounds of stimuli (music) and subjects (annotators), stimulus types and annotation methods are evaluated and the results suggested that cross-cultural and cross-dataset predictions of both valence and arousal values could achieve comparable performance to within-dataset predictions. We also discuss how the generalizability of regression models can be affected by dataset characteristics. Findings of this study may provide valuable insights into music mood regression for nonWestern and other music where training data are scarce.