EmoStack-Net: A Hybrid Feature Learning Framework for EEG-based Emotion Recognition

Abstract

Electroencephalography (EEG) is a key research field in affective computing and brain-computer interface applications for the recognition of emotion. But the non-stationary and complex nature of EEG signals presents a great challenge for emotion classification. In this paper, a novel hybrid feature learning framework for EEG-based emotion recognition is presented, which combines spectral, temporal and statistical EEG descriptors, named EmoStack-Net. EEG signals are transformed into a robust feature representation by extracting Power Spectral Density (PSD), Hjorth parameters, mean, variance and entropy features from EEG recordings. The proposed framework is tested with the DEAP dataset for binary valence classification. Four classifiers, namely SVM, Random Forest (RF), XGBoost and LightGBM are applied to perform comparative analysis. Experimental results show that the proposed hybrid feature representation is effective in improving the classification performance over the conventional PSD only feature. The highest performance was obtained by the Random Forest model with 81.64% accuracy, highlighting the effectiveness of the multi-modal approach by combining the different domains of EEG features for emotion recognition. In addition, the effect of class imbalance handling method using Synthetic Minority Over-sampling Technique (SMOTE) is studied. The proposed framework is a computational efficient solution to EEG based affective computing systems and also gives a foundation for further subject dependent emotion recognition research.