Adaptive graph convolution domain adaptation network for cross-subject EEG emotion recognition

Electroencephalogram (EEG)-based emotion recognition is essential for the advancement of affective brain-computer interface (aBCI) system. However, in cross-subject scenarios, the dynamic nature and subject-specific characteristics of EEG signals significantly hinder knowledge transfer, thereby leading to reduced model performance on previously unseen target domain. To overcome these limitations, we design an innovative domain adaptation model, the adaptive graph convolution domain adaptation network (AGCDAN), to capture the dynamic spatial information of EEG signals and reduce inter-domain distribution discrepancies by aligning both marginal and conditional distributions through domain adaptation. Specifically, adaptive graphs are firstly constructed based on differential entropy features extracted from EEG signals to extract dynamic frequency-spatial representations. A multi-branch neural network is then employed to extract customized feature representations tailored to each source domain and the target domain individually. Subsequently, discriminator-free adversarial learning is employed to align marginal distributions, and introduces subdomain metric learning guided by label information to achieve conditional distribution alignment. Finally, domain-specific classifiers, combined with a decision fusion strategy, produce the final emotion predictions. We evaluate AGCDAN on three datasets (SEED, SEED-IV, DEAP) under a multi-source domain adaptation setting for cross-subject emotion recognition. Achieving recognition accuracies of 89.68%, 68.61%, and 68.13%, respectively, demonstrating superior performance over current state-of-the-art (SOTA) domain adaptations techniques in cross-subject emotion recognition, showcasing its strong capability in modeling dynamic emotional states and reducing negative transfer effects.