Coarse-to-Fine Domain Adaptation for Cross-Subject EEG Emotion Recognition with Contrastive Learning

Electroencephalography (EEG) signals have been reported to be informative and reliable for emotion recognition in recent years. However, the accurate recognition across subjects is still challenging because of the large variability of EEG signals. Inspired by the idea of domain adaptation which aims to transfer knowledge learned from source domain to target domain, we propose a novel coarse-to-fine domain adaptation method based on contrastive learning. In the proposed method, the maximum mean discrepancy metric is first employed to approach the distance of EEG data between source and target domains for global alignment. And then for local alignment, we use local maximum mean discrepancy with contrastive learning to reduce the distance of EEG data with the same emotion label and push apart samples with different emotion labels in different subdomains. Moreover, a strategy of classrelevant sample optimization is also designed to reduce biases caused by different distributions of target data. To verify the effectiveness of our method, we perform the experiments on the SEED and SEED-IV datasets, and achieve the recognition accuracies up to 86.44 +/- 4.22% and 82.81 +/- 5.89% on average respectively. This validates that the proposed coarse-to-fine domain adaptation method can supply a reliable solution for cross-subject emotion recognition.