Abstract: In cross-subject emotion recognition, there are significant differences in the distribution of electroencephalogram (EEG) samples among different subjects, and domain adaptation is commonly used to alleviate the differences. However, the differences in the EEG distribution across affective subdomains are ignored by global adaptation, which reduces the distinguishability of emotional features. Besides, EEG contains a number of electrodes, and subjects only reach the prospective emotion during part of stimuli. Learning the complex spatial information between channels and emphasizing critical EEG frames is essential. Hence, we propose a subdomain adaptation and spatial-temporal learning network for EEG-based emotion recognition. In the subdomain adaptation module, the difference loss in subdomains is reduced by minimizing intra-class differences and maximizing inter-class differences. A spatial-temporal feature extractor captures spatial correlations and temporal contexts, extracting discriminative emotional features. Subject-independent experiments conducted on the public DEAP dataset demonstrate the superior performance of the proposed method, achieving classification accuracies of 0.688 0 for arousal and 0.696 8 for valence, respectively.

Key words: emotion recognition, electroencephalogram (EEG), subdomain adaptation, spatial-temporal learning, deep learning