Hierarchical dynamic local-global-graph representation learning for EEG emotion recognition

초록

Recent advancements in emotion-state measurement using electroencephalography (EEG) have emphasized the importance of modeling the spatial relationship of brain signals. However, most existing GNN-based methods rely on single-level spatial modeling, which overlook the intrinsic spatial characteristics of EEG signals, including high intra-region similarity and distinct inter-region differences. In this work, we propose HDGNet, a hierarchical dynamic local-global graph representation learning method for the measurement of affective states. HDGNet incorporates both local and global graph structures based on prior knowledge to capture brain activities at multiple scales. Through hierarchical GNN modules, HDGNet processes interactions dynamically both within localized regions and across broader brain areas, enabling efficient learning of emotion-relevant patterns. In this paper, we focus on a comparative study on four datasets: SEED, SEED-IV, SEED-V, and DREAMER. The experiment results demonstrate that the proposed method enhanced measurement accuracy and reliability across diverse datasets. Notably, the proposed model achieves accuracy improvements ranging from 6.95% to 14.39% on electrode-rich datasets (SEED, SEED-IV, SEED-V) compared to baseline. Further ablation and visualization analyses validate the effectiveness of HDGNet in capturing both intra- and inter-regional brain dynamics across varying emotional states. The code is available at https://github.com/affctivai/HDGNet. © 1963-2012 IEEE.

키워드