Adaptive Multi-Domain EEG Feature Fusion with Augmented Machine Learning Framework for Four-Class Emotion Recognition
EEG based Emotion recognition has gained a lot of interest in the domain of Affective Computing, Healthcare and Human-Computer Interaction. Emotion classification in EEG signals remains difficult, though, because of their nonlinearity, noise and dependence on the specific person. In this paper, we suggest an Adaptive Multi-Domain EEG Feature Fusion Framework for four-class emotion recognition. The proposed method is a hybrid technique composed of frequency-specific band-pass filtering, denoising using wavelet, multi-domain feature extraction, data augmentation and dimensionality reduction using PCA. Three types of features are extracted: time domain, Fast Fourier Transform (FFT) domain and Power Spectral Density (PSD) domain features, and these features are combined to create a feature representation. Gaussian noise injection, temporal shifting and amplitude scaling are used as augmentation strategies to obtain a better generalization. Optimized feature set is classified with Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Random Forest (RF), Decision Tree (DT), Naïve Bayes (NB) classifiers. The effectiveness of the proposed framework for robust emotion recognition using EEG data has been demonstrated through experiments. The classifiers evaluated and found that the highest classification accuracy was for the Random Forest model with 99.6%, followed by KNN model with 99.5%, SVM model with 99.2%, and Decision Tree model with 99.1%. Naïve Bayes had a relatively poor accuracy rate of 92.0%, however. The findings demonstrate that the proposed multi-domain EEG feature fusion framework is more effective to recognize emotions accurately by four classes.