Minimal Eeg Electrode Selection for Emotion Recognition: A 2-Channel Cross-Subject Study on the Dreamer Dataset

This is the authors' abstract. We don't add key points for this paper.

EEG-based emotion recognition holds considerable promise for affective computing, neuromarketing, and brain-computer interface (BCI) applications, yet its practical deployment remains constrained by the requirement for dense electrode configurations. This study investigates whether a minimal subset of electrodes from the consumer-grade Emotiv Epoc headset (14 channels) can preserve or exceed emotion classification accuracy, enabling low-cost wearable emotion recognition. The DREAMER multimodal affective dataset (23 subjects, 14-channel EEG, 18 video stimuli) was used, and log band power features were extracted across five frequency bands: theta, alpha, beta-low, beta-high, and gamma. Three complementary electrode importance ranking methods — mRMR, SHAP, and permutation importance — were applied within a Leave-One-Subject-Out (LOSO) cross-validation loop to produce a leakage-free consensus electrode ranking. A systematic ablation study evaluated SVM (RBF kernel) and LDA (SVD solver) classifiers for top-ranked subsets ranging from 2 to 14 channels. The top-ranked 2-channel subset (F4 and P7) achieved 62.39% valence and 75.43% arousal accuracy with SVM, exceeding the full 14-channel baseline (53.36% and 74.62% respectively) by +9.03 percentage points for valence. This improvement was statistically significant (Wilcoxon signed-rank test: W = 266.0, p < 0.001; paired t-test: t = 4.477, p < 0.001), with 20 of 23 subjects individually showing improved performance. Notably, LDA maintained consistently higher valence accuracy (58.99%–61.94%) than SVM across all subset sizes, with the gap widening at larger subsets, corroborating a bias-variance overfitting interpretation. An 86% reduction in electrode count was achieved with no loss in classification performance. The arousal-valence accuracy gap of 21.26% observed at baseline narrowed to 13.04% under the minimal electrode configuration, consistent with neurophysiological theory.