A machine learning approach to EEG based prediction of user's music preferences

Abstract: Music has many benefits for our mood and feelings, especially so when we get to choose our own favorite music. However, accessing one's favorite music is not as easy for everyone. For motorically disabled and locked-in people, interacting with devices used for listening to music is challenging since it requires physical interaction. Machine learning classification methods used with EEG could prove useful for detecting individual musical preferences, extracted without any physical or verbal interaction. The two most common methods within EEG-based classification are Artificial neural networks (ANN) and Support vector machines (SVM). This study compares the performance of these two methods, when used with the DEAP dataset of EEG-monitored participants watching music videos. This comparison can help with gaining insight into which machine learning method is most appropriate for music preference detection, contributing towards more accurate predictions of motorically disabled people’s musical preferences. The participants in the DEAP dataset rated watched music videos by preference from 1 to 9, of which we trained the classification models to separate between higher (chosen as ratings 8 to 9) and lower ratings. From the results it is concluded that ANN performs better than SVM in terms of accuracy, with ANN performing at roughly 86% and SVM at 85%, while the SVMs were substantially faster to train. These accuracy scores were obtained from two ANN and SVM models using the optimal parameter and channel configurations, which were calculated through extensive testing. The accuracies are however likely achieved due to an imbalanced dataset, with too few data samples of higher ratings in proportion to lower, leading to biased classifiers that work well on our dataset but has probably close to random classification performance. Alterations to our methods could give better performing classifiers, and would also lead to more meaningful comparisons of ANN and SVM for EEG-based musical preference prediction.