Geometric Deep Learning for the Prediction of Music Liking from Consumer-Grade EEG

Abstract

Abstract Music evokes meaningful experiences in almost everyone, but musical tastes are highly dependent on the individual. By making use of a large dataset of EEG recordings during music listening and associated participant music ratings, the performance of modern deep neural network models for predicting music liking from EEG recordings using temporal, spectral and connectivity-based features is explored. An end-to-end trainable graph neural network based on sinc bandpass lters and functional connectivity is proposed and successfully tested, using am- plitude envelope correlations as a di erentiable connectivity measure. Results show that subtracting the resting state connectivity from the connectiv- ity during music listening serves as an e ective baselining procedure when training to predict music ratings across subjects. When using personalised models to pre- dict music ratings, there is a tendency for di erent models to perform well on the same subjects, which is not explained by the investigated subject-related features. A sex classi cation task was used to validate the dataset and model implemen- tation, which revealed that sex classi cation performance is substantially higher during music listening than during resting state.