Graph Learning for Cross-Subject, Cross-Population EEG Emotion Decoding and Model-Derived Spatial-Spectral Neural Signatures

Abstract:Electroencephalography (EEG) provides a noninvasive means of capturing emotion-related neural dynamics, yet reliable EEG emotion decoding lacks models that can both generalize to unseen individuals and populations while preserving neural interpretability. To address these challenges, EmoDiPyraTrans is proposed as a development-regularized differential graph Transformer that models temporally ordered relative power spectral density graphs through adaptive graph recurrence, differential attention, and multiscale fusion. The framework was evaluated at three connected levels. First, cross-subject evaluations on SEED, FACED, MAHNOB-HCI, DEAP and DREAMER yielded participant-mean accuracies of 0.928, 0.645, 0.714, 0.617 and 0.671, respectively; the model ranked first among the evaluated methods for accuracy and positive-class F1 on all five datasets. Across seven ablation protocols, differential attention was the only component whose removal reduced both metrics in every case, whereas removing maximum mean discrepancy reduced accuracy throughout. Second, DEP-EEG distinguished within- from cross-population positive-versus-neutral decoding. Accuracy was $0.802$ within healthy controls, 0.704 within participants with depression and 0.591 under healthy-to-depression transfer. Mixed-population development produced $0.581$ accuracy and the highest positive-class F1 (0.498), indicating that greater population diversity alone did not remove the transfer gap. Third, channel- and frequency-resolved analyses on SEED identified a distributed frontal, temporal, central and parietal pattern, an alpha-centred low-to-mid-frequency preference and a six-channel subset that preserved near-full performance.

Signal Processing (eess.SP); Machine Learning (cs.LG)

arXiv:2609.22103 [eess.SP]

(or arXiv:2609.22103v1 [eess.SP] for this version)

https://doi.org/10.48550/arXiv.2609.22103 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.