Ensemble Machine Learning-Based Affective Computing for Emotion Recognition Using Dual-Decomposed EEG Signals

Machine learning (ML)-based algorithms have shown promising results in electroencephalogram (EEG)-based emotion recognition. This study compares five ensemble learning-based ML (EML) algorithms with five conventional ML (CML) algorithms for recognizing multiple human emotions from EEG signals. A publicly available DREAMER database having nine emotions is used to design ML-based system, which is validated on SEED, INTERFACES, and MUSEC databases. In this study, initially, EEG signals are separated into theta, alpha, beta, and gamma bands by applying discrete wavelet transform and then empirical mode decomposition is applied for further decomposition of band-separated EEG signals into intrinsic mode functions (IMFs). Then, 31 statistical features are extracted from IMFs to design ML-based system using five multiclass EML algorithms such as bagging, random forest, rotation random forest, extreme gradient boost, and adaptive boosting. Finally, the performance of these five EML algorithms is evaluated using 10-fold cross-validation and compared against five CML algorithms using performance evaluation metrics such as accuracy, F1-score, kappa-score, and area-under-the-curve (AUC). The mean accuracy of multiclass emotion recognition over five EML algorithms is ~5.87% and ~6.08% higher than the mean accuracy of five CML algorithms, for both arousal (88.95% vs. 83.08%) and valence (88.90% vs. 82.81%) dimensions, respectively. The EML-based bagging algorithm reported the highest accuracy, F1-score, kappa-score, and AUC of 95.81%, 0.81, 0.79, and 0.98, respectively for arousal and 95.53%, 0.80, 0.77, and 0.98, respectively for valence. A similar trend is also observed on the three validation datasets. The EML algorithms provide better multiclass emotion recognition compared to CML algorithms.