Enhancing Cross-Session EEG Emotion Recognition Through Multi-Source Domain Adaptation and Association Reinforcement

Emotion recognition from electroencephalogram (EEG) signals is essential for numerous applications but presents significant challenges due to the variability between subjects and the non-stationary characteristics of EEG data. Our method effectively addresses the discrepanices in conditional distributions of EEG signals across different sessions and subjects. Key innovations include the use of multi-source domain adaptation to process multiple related but distinct data distributions, a novel association loss calculation method to capture both inter-domain relationships and intra-domain correlations, and the integration of common and domain-specific encoders for efficient feature extraction. These components collectively enhance the model's capacity to manage cross-session variations in EEG data. We evaluate the approach using two public datasets: SEED and SEED-IV. The proposed model sets a new benchmark in performance, achieving accuracies of 89.38% on SEED and 66.17% on SEED-IV, surpassing existing methods. The results highlight the effectiveness of our approach in addressing the non-stationarity issues inherent in EEG signals, leading to improved emotion recognition accuracy across different sessions and subjects. This research advances the development of more robust and adaptable EEG-based emotion recognition systems, with potential applications in healthcare, education, and human-computer interaction.