A Novel Emotion Recognition Method Incorporating MST-based Brain Network and FVMD-GAMPE

Emotion recognition is a key technique of intelligent human-computer interaction (HCI) systems. In the current research on emotion recognition, there are several limitations such as inconsistent brain network scale and high time complexity of modal decomposition. To overcome these shortcomings, we propose a novel emotion recognition method incorporating MST-based brain network and FVMD-GAMPE. Firstly, electroencephalography (EEG) data is decomposed into four frequency bands $(\theta,\alpha,\beta,\gamma)$ by wavelet packet transform (WPT), and mutual information (MI) between channel pairs is calculated to construct the connectivity matrix. Secondly, the brain network based on the minimum spanning tree (MST) is constructed and seven features are extracted. Thirdly, fast variational modal decomposition (FVMD) and WPT are applied to process EEG data to obtain the variational mode functions (VMF) of different frequency bands. Then, the parameters of the multi-scale permutation entropy (MPE) are optimized with the genetic algorithm (GA), and then MPE features are extracted. Finally, the features extracted from MST-based brain network are fused with MPE features, and then fused features are fed to the random forest (RF) classifier to recognize emotional states. Experimental results on DEAP show that the best classification accuracy for valance and arousal are 89.58% and 88.54%, respectively. The result analysis demonstrates MST-based brain network in the negative emotional states has a more divergent topology. This means that brain regions are more active and have a faster exchange of information flow when the brain processes negative emotions. On the other hand, brain network of women is similar to a star-shaped structure, which indicates women’s brain activation is higher than man. This study provides theoretical support for research on negative bias.