SS-MSDA: Streamlined Sample-level Multi-source Domain Adaptation for EEG Emotion Recognition

Affective EEG-based Brain-Computer Interface (BCI) offers extensive prospects. Yet, it grapples with notable challenges in consistently achieving accurate emotion recognition among new subjects. Mitigating this matter, Multi-Source Domain Adaptation (MSDA) has been advanced. However, they exhibit performance that falls short of expectations, necessitate complex preparations and lack solid theoretical underpinnings. Concerning these insufficiencies, we propose an innovative MSDA algorithm, effectively narrowing the Wasserstein Distance between identified subdomain and the target domain, thereby theoretically constraining the upper bound of emotion classification error. Compared with baseline model on the emotional EEG dataset SEED,SS-MSDA achieved an increase in recognition accuracy ranging from 1^~14% (average improvement of 7.2%) across subjects, demonstrating superior performance over Domain Adaption (DA) benchmarks. Moreover, it significantly reduced the preparation time by over 99.8%, along with its minimal computational costs, thus being exceptionally apt for practical applications. Finally, the algorithm extends the theory of MSDA for affective BCI. The significance of this algorithm lies in its potential to improve the recognition accuracy of existing emotion recognition algorithms on new subjects, without the need for pre-training and data pre-collection. Moreover, it provides a novel theoretical perspective on the methods for constraining the error upper bound of the classifier.