MG-DA: Multi-branch graph convolution with domain adaptation for emotion recognition

Emotion recognition based on electroencephalography (EEG) is a critical task in the field of affective computing. However, the non-stationarity and individual differences in EEG signals pose significant challenges in developing a model with high generalizability. Although current approaches attempt to address this issue using domain adaptation techniques, the similarity in neural patterns across different emotional states often makes it difficult to accurately capture fine-grained differences in features. In this study, we propose an efficient Multi-Branch Graph Convolution with Domain Adaptation (MG-DA) for EEG topology-adaptive calibration. MG-DA is consisted of two key steps: a multi-branch graph convolution module and domain adaptation strategy. The multi-branch graph convolution comprehensively extract topological features across EEG frequency bands. In domain adaptation strategy, three crucial modules are designed to enhance the alignment of source and target domain features, including a spatial distribution module for constraining the spatial distribution of relevant features, domain alignment for reducing the domain difference, and class confusion module for enhancing the boundaries among classes. We conducted cross-subject and cross-session experiments on three publicly available datasets, SEED, SEED-IV and DEAP. The experimental results demonstrate that our model outperforms other state-of-the-art methods in terms of generalization ability.