EEG emotion recognition using a channel-aware feature encoder and XGBoost classifier

Electroencephalogram (EEG)-based emotion recognition is hindered by a low signal-to-noise ratio, inter-subject variability, and the computational cost of deep models. We propose MAFNet-XGBoost, a high-efficiency hybrid framework where differential entropy and power spectral density features are first mapped by a fully connected neural network (FCNN) and then encoded by a dual-branch Adaptive Channel-wise Feature Encoder (ACFE) to model global inter-channel dependencies and local temporal dynamics. The fused representations are classified by eXtreme Gradient Boosting (XGBoost). On SEED-IV and DEAP, MAFNet-XGBoost achieves 85.64% and 94.15% accuracy, outperforming existing methods. Granger causality analysis supports physiological consistency and interpretability.