Accurate an‎d robust recognition of human emotional states is essential for building harmonious human–machine interaction systems an‎d advancing brain–computer interface (BCI) technologies. Electroencephalography (EEG), owing to its non-invasive nature, low cost, an‎d high temporal resolution, provides a powerful physiological signal for decoding emotions by capturing subtle neural patterns directly linked to brain activity. However, cross-session variability in EEG data poses a major challenge to model generalization. To address this, we propose EEG-based Graph-Guided Domain Adaptation (EGDA), a novel framework that jointly aligns both marginal an‎d conditional distributions across domains while preserving the intrinsic geometric an‎d semantic structures of EEG data through graph regularization. Furthermore, we employ a pseudo-target label refinement mechanism that iteratively converges toward a stable solution, enhancing the robustness of domain alignment. By formulating a tailored optimization problem, EGDA achieves a more effective feature adaptation an‎d transfer learning strategy. Experiments conducted on the SEED-IV dataset demonstrate that the proposed approach significantly improves recognition accuracy an‎d consistently outperforms baseline methods. These findings highlight the potential of graph-guided domain adaptation as a promising direction for robust EEG-based emotion recognition an‎d intelligent human emotion analysis.

حامد نريماني , فرزانه شايق بروجني

جلال ذهبي , امير اخوان بي تقصير