Learning from Brain Topography: A Hierarchical Local-Global Graph-Transformer Network for EEG Emotion Recognition

TL;DR

This work tackles EEG-based emotion recognition by leveraging brain topography to model both local electrode topology and global brain dynamics. It introduces Neuro-HGLN, a two-stream architecture that combines a spatially grounded Global Graph Learning Stream with a Hierarchical Local-Region Stream, where region-specific graphs are learned and integrated by an iTransformer using a dimension-as-token approach; the model is regularized by a Geometric Constraint KL divergence and a Functional Diversity term. Core contributions include a Spatial Euclidean Prior $\\mathbf{A}_{prior,ij} = \\exp(-\\mathbf{D}_{ij}^2 / \\tau)$, dynamic fusion to obtain $\\mathbf{A}_g$, region-wise learned adjacencies $\\mathbf{A}_{local}^k$, and an end-to-end loss $\\mathcal{L}_{total} = \\alpha\\mathcal{L}_{dist} + \\beta\\mathcal{L}_{div} + \\gamma\\mathcal{L}_{global} + \\delta\\mathcal{L}_{local}$. Empirically, Neuro-HGLN achieves state-of-the-art results on SEED, SEED-IV, SEED-V, and MPED in both subject-dependent and subject-independent settings, with improved interpretability tied to neurophysiological structure. The framework demonstrates that unifying local topological learning with cross-region dependencies enhances robustness and fine-grained emotion recognition in EEG signals, offering a practical impact for affective computing and brain-computer interfacing.

Abstract

Understanding how local neurophysiological patterns interact with global brain dynamics is essential for decoding human emotions from EEG signals. However, existing deep learning approaches often overlook the brain's intrinsic spatial organization, failing to simultaneously capture local topological relations and global dependencies. To address these challenges, we propose Neuro-HGLN, a Neurologically-informed Hierarchical Graph-Transformer Learning Network that integrates biologically grounded priors with hierarchical representation learning. Neuro-HGLN first constructs a spatial Euclidean prior graph based on physical electrode distances to serve as an anatomically grounded inductive bias. A learnable global dynamic graph is then introduced to model functional connectivity across the entire brain. In parallel, to capture fine-grained regional dependencies, Neuro-HGLN builds region-level local graphs using a multi-head self-attention mechanism. These graphs are processed synchronously through local-constrained parallel GCN layers to produce region-specific representations. Subsequently, an iTransformer encoder aggregates these features to capture cross-region dependencies under a dimension-as-token formulation. Extensive experiments demonstrate that Neuro-HGLN achieves state-of-the-art performance on multiple benchmarks, providing enhanced interpretability grounded in neurophysiological structure. These results highlight the efficacy of unifying local topological learning with cross-region dependency modeling for robust EEG emotion recognition.

Figures (3)

• Figure 1: The overall architecture of the proposed Neuro-HGLN framework. The model orchestrates two synergistic streams to process EEG DE features: (a) The Global Graph Learning Stream captures holistic brain dynamics by integrating a static Spatial Euclidean Prior with learnable dynamic weights to construct a global graph, which is then processed by stacked GCN layers. (b) The Hierarchical Local-Region Stream models fine-grained dependencies by partitioning the brain into anatomical regions. It employs Local-constrained Parallel GCNs to extract intra-regional topology, followed by an iTransformer Encoder that captures high-level cross-region semantic dependencies using a dimension-as-token mechanism. The final emotion prediction is achieved by fusing the logits from both streams.
• Figure 2: Visualization of learned topological structures. (a) The Global Graph captures holistic, long-range dependencies. (b-f) The Local Graphs exhibit distinct, region-specific activation patterns (e.g., Prefrontal focus vs. Occipital focus), verifying that the Functional Diversity Regularization effectively prevents feature homogenization.
• Figure 3: t-SNE visualization of feature distributions on the MPED dataset. The digit labels 0--6 correspond to neutrality, joy, funny, anger, fear, disgust, and sadness, respectively. (a) PGCN shows a scattered distribution with severe class overlap, indicating poor separability. (b) Neuro-HGLN forms distinct, compact clusters with clear inter-class margins, demonstrating its superior ability to disentangle fine-grained emotional features.