An Attention-Enhanced Retentive Broad Learning System for Subject-Generic Emotion Recognition from EEG Signals

Abstract

Emotion recognition (ER) utilizing electroencephalography (EEG) is significant in affective brain-computer interface research. Recent advances have underscored the supremacy of deep learning-based ER techniques over traditional statistical methods. Still, challenges persist in extracting subject-specific and subject-shared features across temporal, spatial, and frequency domains for transferable EEG-based ER. We propose an attention-enhanced naïve-gated broad learning system (ANGB) to tackle these issues. It includes a causality-based dual-routing attention encoder that uncovers dynamic affective process aspects by integrating band dependence and channel coupling. Moreover, it incorporates a naïve gated recurrent unit within the broad learning system, modeling complex inter-source relationships and proficiently acquiring domain-specific and domain-shared functionalities. Extensive experiments on the DEAP and MAHNOB-HCI databases demonstrate the commendable performance of our proposed model in the context of subject-generic ER.