A Hierarchical Three-Dimensional MLP-Based Model for EEG Emotion Recognition

Abstract

Electroencephalogram (EEG) sensor data are useful and important for emotion recognition. However, cross-subject EEG emotion recognition suffers from the challenging problems of individual difference and noise disturbance. To cope with these problems, we propose a hierarchical 3-D MLP-based neural network (HMNN). This method consists of multiple hierarchical layers of 3D-MLPBlocks and a noise optimization module. The 3D-MLPBlock is designed to extract the multiperiod features of common emotional patterns across different individuals; the noise optimization module is devised to enhance the network robustness to noise disturbance. Experimental results on public benchmarks DEAP, DREAMER, and SEED-IV have demonstrated the superiority of HMNN over the related advanced approaches. Specifically, HMNN obtains the accuracies of 63.69%/60.03% for valence/aoursal classification on DEAP, 62.51%/64.49% for valence/arousal classification on DREAMER, and 62.29% for emotion classification on SEED-IV.