A Knowledge Transfer Method from Multi-Channel to Single-Channel EEG Toward Real-World Emotion Recognition Deployment

Multi-channel electroencephalography (EEG) provides spatial information for emotion recognition but requires more complex signal acquisition than single-channel EEG. This study proposes a multi-channel-to-single-channel knowledge transfer framework that uses multi-channel recordings during training and one selected physical EEG channel at inference. A 32-channel teacher is pretrained, and Sparsemax-based gating is jointly optimized with a student using classification, logits distillation, frequency-band consistency, and gating regularization losses. The frequency-band objective compares signal-level spectral statistics of the multi-channel recordings and the gated surrogate input. After channel locking, the student is retrained on the selected physical signal. Binary Arousal and Valence classification was evaluated on DEAP using 32-fold leave-one-subject-out cross-validation. The proposed achieved higher subject-averaged balanced accuracy and F1-score than Channel-matched CE in both tasks, although component effects varied across tasks and metrics. These findings support the evaluated training approach under single-channel input constraints and provide a methodological basis for simplifying EEG acquisition.