HRGNN: Hierarchical region-aware graph neural network for interpretable EEG-based emotion recognition.

Abstract

Neurophysiological studies have shown that cortical information processing involves complex interactions among multiple functional brain regions. However, most existing EEG-based emotion recognition methods employ Graph Neural Networks (GNNs) to model connectivity between electrodes, often neglecting functional specialization and local spatial dependencies across brain regions. This leads to insufficient granularity in feature representation. In addition, conventional GNNs frequently suffer from over-smoothing, while global average pooling strategies tend to overlook information from critical regions. To address these challenges, we propose a Hierarchical Region-Aware Graph Neural Network (HRGNN). The model first incorporates a Brain Region Embedding (BRE) module, which leverages topological priors of brain region partitioning to map the whole-brain EEG signals into structured multi-region subgraph representations. Then, a region-aware graph encoder is introduced, combining regional aggregation and hierarchical graph pooling to capture multi-scale intra- and inter-regional interactions effectively. To further enhance discriminative representation, a Dynamic Routing-based Mixture-of-Experts (DMoE) module is developed to fuse regional subgraph features adaptively, assigning greater importance to brain regions with higher emotional relevance. Extensive experiments on eight publicly available EEG emotion datasets demonstrate that HRGNN consistently outperforms state-of-the-art methods. Visualization analyses further validate the model's effectiveness in emotion discrimination and its alignment with neuroscientific principles of brain region interaction.