TOWARDS MINIMAL EEG FOR CROSS-SUBJECT EMOTION RECOGNITION: A STUDY ON THE SEED DATASET

Abstract

Affective brain-computer interfaces (BCIs) based on electroencephalography (EEG) typically require research-grade, full-cap electrode setups that are impractical for wearable and real-world deployment. This paper investigates the minimum number of electrodes needed to perform dependable three-class emotion recognition (Negative, Neutral, Positive) on the SEED dataset, a widely used benchmark comprising 15 subjects with 62-channel EEG recordings. Differential Entropy (DE) features are extracted across five frequency bands (delta, theta, alpha, beta, gamma) and a consensus electrode ranking is derived via a novel nested Leave-One-Subject-Out (LOSO) procedure combining minimum Redundancy Maximum Relevance (mRMR), SHAP-based Random Forest importance, and permutation importance — ensuring zero data leakage between ranking and evaluation. An ablation study across 13 electrode-count configurations reveals that an 8-electrode subset (T7, FT8, TP7, T8, FT7, C5, F8, FC5), all located in the bilateral temporal and fronto-temporal regions, achieves 61.45% ± 8.48% LOSO-CV accuracy — a statistically significant improvement of +12.60 percentage points over the full 62-channel baseline of 48.85% ± 16.28% (Wilcoxon signed-rank: W = 11.0, p = 0.0034, N = 15). Critically, the standard deviation of per-subject accuracy is nearly halved (16.28% → 8.48%), indicating substantially improved cross-subject consistency. These findings demonstrate that temporal-region electrodes encode the dominant emotion-discriminative signal in SEED stimuli, and that compact electrode configurations can outperform full-cap systems under cross-subject evaluation — with direct implications for wearable affective computing.