MMPF-Net integrates temporal, spectral and spatial projection features for EEG emotion recognition

Electroencephalography (EEG) emotion recognition is challenged by non-stationarity and distribution shifts across sessions and participants. We propose MMPF-Net, a parallel architecture that extracts temporal, STFT-based spectral, and learned spatial-projection representations from the same EEG segment. Three 128-dimensional branch features are aligned by L2 normalisation and combined through sample-wise gated concatenation. Evaluation included trial-level within-subject five-fold cross-validation, leave-one-session-out, and leave-one-subject-out protocols on SEED, plus external within-subject validation on DEAP. MMPF-Net achieved 96.53 ± 0.82%, 85.76 ± 4.96%, and 70.25 ± 5.94% Accuracy under the three SEED protocols, respectively, and 90.03 ± 4.51% and 89.26 ± 4.73% for DEAP valence and arousal. Participant-level paired analyses showed statistically supported Accuracy gains over EEG Conformer across the reported protocols. Extended ablations indicated contributions from all three branches, sample-wise gating, and the frequency branch beyond parameter count, while several alternative weighted-scoring schemes were not significantly different after Holm correction. These findings support MMPF-Net for participant-specific EEG emotion decoding, while performance under session, participant, and dataset shifts remains limited.