Functional Connectivity Patterns Learning for EEG-Based Emotion Recognition

Abstract

Neuroscience research reveals that different emotions are associated with different functional connectivity structures of brain regions. However, many existing electroencephalography (EEG)-based emotion recognition methods use these connectivity patterns broadly without distinguishing between specific emotions. Additionally, the nonstationarity of EEG signals often results in high variations across different periods, leading models to extract time-specific features instead of emotional features. This article proposes a functional connectivity patterns learning network (FCPL) for EEG-based emotion recognition to address these challenges. FCPL includes a coefficient branch, a graph construction module, and a period domain adversarial module. These components capture individual characteristics and specific emotional connectivity patterns and reduce period-related variations, respectively. FCPL achieves state-of-the-art results: 42.04%/28.81% for seven-class subject-dependent/independent experiments on the MPED dataset, 97.45%/89.88% for subject-dependent/independent experiments on the SEED dataset, and 95.98%/96.19% for valence/arousal subject-dependent experiments and 67.90%/65.60% for valence/arousal subject-independent experiments on the DREAMER dataset. This work advances the exploration of functional connectivity structures in EEG signals from coarse-grained emotion-related patterns to fine-grained emotional distinctions, promoting neuroscience, and EEG-based emotion recognition technologies.