Feature Transfer Learning in EEG-based Emotion Recognition

This is a challenge to enhance the distribution between the source and the target domains between different subjects, which is difficult but important for practical applications. In this paper, we proposed a transfer framework based on feature analysis for EEG-based emotion recognition. We firstly extract differential entropy features by computing logarithms of power spectral density. On this basis, transfer component analysis (TCA) was used to project the source and the target domains into a kernel Hilbert space, in order to reduce the distance between the two domains. Finally, we compare the performance of emotion recognition among different dimensions. Experiments results show that the best mean accuracy is 58.49% by using TCA-based method, which is better than the previous study of 52.06%. Meanwhile, experimental results validate the feasibility and efficiency for subject transfer emotion recognition.