Emotion Specific Network with Multi-dimension Features in Emotion Recognition

Abstract

Emotion recognition has been recognized as an important issue in terms of human-computer interaction. Various studies showed brain regional cooperation changed with mental state. However, the important role of specific brain channels and their topology during the emotion activity is still unclear. In this paper, we extracted the multi-dimension EEG features to achieve emotion specific network construction and emotion recognition. The dataset is from the 2020 World Robot Conference-Brain-Computer Interfaces (BCI) Contest, provided by Shanghai Jiaotong University. There are 24 sessions included in this paper. The power spectrum density (PSD), Hjorth parameter, and functional connectivity were extracted from each session. A data-driven critical channel selection strategy was performed by sorting the classification accuracy of each channel. Meanwhile, the emotion specific network was established by the top 10 channels. Finally, we mixed the features in the emotion specific network to recognize three emotion types (positive, neutral, and negative). We found the reorganization of the brain network mostly focused on the right hemisphere. Moreover, the classification accuracy (70.53% ± 4.61% (mean ± std)) showed the feasibility of emotion specific network. Our study indicated the emotion specific network is critical for emotion alteration and provides a new insight for emotion recognition.