Native relation learning for cross-dataset EEG emotion recognition
Biomedical Signal Processing and Control · 2027-02 · 130 · 111581
DOI: 10.1016/j.bspc.2026.111581 ↗Kurz gefasst
Fachartikel in der Zeitschrift Biomedical Signal Processing and Control, 2027-02. Verfasst von Zhenwei Cao und 4 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Gefördert von: Graduate Research and Innovation Projects of Jiangsu Province, National Natural Science Foundation of China.
Abstract (Originalsprache)
Electroencephalography (EEG) emotion recognition is widely used in affective computing, brain–computer interfaces, and mental-state assessment. Its performance, however, often drops when a model is transferred across subjects or datasets because EEG recordings vary with individual neural responses, elicitation paradigms, acquisition settings, and emotion-label definitions. Public EEG emotion datasets also differ in the number of categories and in how labels are obtained. Collapsing these datasets into a single output space may simplify training, but it can also remove the native category relations that define the decision boundaries of each target dataset. This study proposes NativeRelation, a native relation learning framework for cross-dataset EEG emotion recognition. During multi-dataset joint pre-training, NativeRelation keeps each dataset in its own emotion label space while sharing an EEG encoder across datasets. Dataset-specific prediction heads, class prototypes, native relation geometry constraints, and subject-level few-shot adaptation are used to shape frozen representations that can be read out from a small number of labeled target-subject samples. Experiments are conducted under a leave-one-dataset-out, subject-independent protocol on five public EEG emotion datasets: SEED, SEED-IV, SEED-V, SEED-VII, and DEAP. NativeRelation achieves 79.89%, 55.74%, 66.13%, 48.42%, and 72.47% accuracy on the five target datasets, respectively, with a macro-average accuracy of 64.53%. Under the evaluated protocol, the results suggest that preserving native emotion relations supports target readout under heterogeneous label spaces and yields a more traceable representation structure for EEG emotion recognition.
Förderung
- Graduate Research and Innovation Projects of Jiangsu Province
- National Natural Science Foundation of China
Zitieren
APA
Cao, Z., Lu, Y., Di, X., Xiao, T., & Zhou, W. (2027). Native relation learning for cross-dataset EEG emotion recognition. Biomedical Signal Processing and Control, 130, 111581. https://doi.org/10.1016/j.bspc.2026.111581
Vancouver
Cao Z, Lu Y, Di X, Xiao T, Zhou W. Native relation learning for cross-dataset EEG emotion recognition. Biomedical Signal Processing and Control. 2027;130:111581. doi:10.1016/j.bspc.2026.111581
BibTeX
@article{Cao2027,
author = {Cao, Zhenwei and Lu, Yonghua and Di, Xinyu and Xiao, Tianxing and Zhou, Weiguo},
title = {Native relation learning for cross-dataset EEG emotion recognition},
journal = {Biomedical Signal Processing and Control},
year = {2027},
volume = {130},
pages = {111581},
doi = {10.1016/j.bspc.2026.111581}
}
EUVIMED