Abstract
Recent advancements in electroencephalography (EEG)-based affective computing have enabled the development of emotion recognition models. However, many of these models require subject-specific training and often fail to generalize across individuals due to variability in brain signals. This paper introduces a novel hybrid model that combines convolutional layers with Transformer-based deep learning to enhance cross-subject emotion recognition without the need for subject-specific calibration. The model is evaluated on the SEED-IV dataset, which includes EEG and eye-tracking data recorded during emotional stimulation. Compared to baseline methods and domain adaptation techniques, the proposed approach achieves significantly higher accuracy and robustness. The findings highlight the model’s potential for real-time, subject-independent emotion recognition in intelligent systems. Limitations include evaluation on a single dataset and focus on only four basic emotions, which will be addressed in future work through broader validation and emotion set expansion.