EEG-based Emotion Recognition Using Multi-scale Window Deep Forest

With the fast development of human-machine interface technology, emotion recognition has attracted more and more attentions in recent years. Compared to other physiological experimental signals frequently used in emotion recognition, EEG signals are easy to record but not easy to disguise. However, because of high dimensionality of EEG data and the diversity of human emotions, feature extraction and classification of EEG signals are still difficult. In this paper, we propose deep forest with multi-scale window (MSWDF) to identify EEG emotions. Deep Forest is an integrated method of decision trees. In the MSWDF, features can be extracted by multi-granularity scanning with multi-scale windows. Compared with deep neural network, the MSWDF not only has less parameters to adjust, but also can realize the classification of the dataset with small samples. In the MSWDF, raw EEG signals were firstly filtered and segmented into samples. Regarding EEG signals as multivariate time series, a new multi-granularity scanning strategy with variable windows is proposed to extract features from EEG samples. After classifying EEG features by the cascade forest, the recognition results are compared with these of Nearest Neighbor algorithm (KNN), Naive Bayes, Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM). We found that the average classification accuracy of three emotions reaches to 84.90%, which is better than those of five compared methods.