Emotion Recognition Based on Baseline-corrected Differential Entropy Features and Sparse Graph Regularized Extreme Learning Machine

In recent years, with the rapid development of brain science and information science, affective Brain Computer Interfaces (aBCI) have become a hot topic. EEG signals which record the electrical activity of neurons in the brain can provide direct evidence for emotional states. Compared with end-to-end deep learning methods, expert knowledge based emotion recognition methods, using hand-crafted features, are computationally cheaper and can have better generalization ability and interpretability. Due to the effort of balancing the energy spectrum of EEG signals, differential entropy features have been proved a powerful tool for emotion recognition. However, differential entropy of EEG signals still exhibits a decrease trend with increasing frequency, resulting in a low representation of the effect of high-frequency components in emotion recognition and do not perfectly solve the problem caused by the 1/f power law of EEG signals. In this paper, a decibel (dB) feature is extracted by introducing baseline normalization of EEG power spectral density (PSD) , which balances the effect of different rhythmic components better. On this basis, for the features are smoothed and a sparse graph regularized extreme learning machine (SGELM) is used to compare the dB feature with DE in emotion recognition. The proposed method is validated on DEAP data set, and the average classification accuracy increased by 11.33% over those of DE features with wild feature smoothing, and by 2.96% with high feature smoothness. The baseline normalized DE feature not only provide a new option for low-cost affective brain-computer interfaces can but also be used as an alternative input feature for deep learning models.