Estimating Music Listener's Emotion from Bio-signals with EEG Denoising and for Unlearned Music Pieces

We have been studying to utilize bio-signals of music listeners for estimating their emotions in order to realize a music recommender system. Our previous study shows high classification accuracy of emotions by applying CNN to bio-signals including brainwave, heartbeat, and pupil diameter. However, there are three remaining issues of small dataset, noise in brainwave, and emotion estimation for unlearned music pieces. Therefore, in this paper, by applying CNN and Random Forest, we compare the emotion classification accuracy with and without brainwave denoising, and analyze the accuracy for unlearned music pieces, where 100 music pieces are used in the experiment for one subject. The comparison results show that the denoising increases the classification accuracy, while the unlearned music pieces are not well classified with slightly higher accuracy than the chance level, which remains for the further study.