An optimal transport-based domain adaptation method for enhancing cross-subject EEG emotion recognition

Electroencephalography has become a very popular, non-invasive method for brain signal measurement. However, it suffers from large inter-subject variabilities, which makes it challenging to perform analysis with good generalization. This work examines emotion recognition from EEG signals in a cross-subject setting. We identify this task as a domain adaptation problem and describe a method that combines feature engineering, the Wasserstein Barycenter Transport algorithm, and a support vector classifier to perform classification in this setting. This novel approach is fully unsupervised, based on optimal transportation, and achieves state-of-the-art results on the SEED-IV benchmark dataset under a leave-one-subject-out protocol.