acceptodds
Under review as a conference paper at ICLR 2027

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

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 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.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG emotion recognition, Graph Neural Networks, Brain Topography, Hierarchical Representation Learning, Graph-Transformer

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KGYdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Neuro-HGLN, a hierarchical graph–Transformer framework for EEG-based emotion recognition. The method consists of two complementary streams. The global stream combines a spatial Euclidean prior derived from electrode positions with a learnable adjacency matrix to capture global dependencies. The local stream constructs multiple attention-based graphs, processes them through parallel GCNs, and employs an iTransformer encoder to model cross-electrode interactions. Geometric consistency and functional diversity regularization are introduced to guide graph learning, and predictions from the two streams are combined through late fusion.

Experiments on SEED, SEED-IV, SEED-V, and MPED report improvements over the selected baselines under both subject-dependent and subject-independent evaluation protocols. Ablation studies and topology visualizations are provided to support the effectiveness of the proposed components.

Critical strengths

The paper presents a coherent dual-stream architecture that combines global graph learning with multiple attention-based graph representations. The joint modeling of global and potentially specialized connectivity patterns is well motivated for EEG emotion recognition and provides a useful architectural direction.

The method achieves competitive results on four EEG emotion recognition benchmarks, including improvements under leave-one-subject-out evaluation. The ablation experiments also suggest that the two streams contribute complementary information to the classification task. These results provide some empirical support for the proposed architecture.

Critical weaknesses

Although the method emphasizes region-level local topology learning, each local graph is defined as a full electrode-to-electrode adjacency matrix without explicit regional constraints. The functional diversity regularization discourages overlapping connections but does not necessarily enforce spatial localization or anatomically meaningful specialization.

Furthermore, the unconstrained learnable projections and biases in the graph construction can substantially transform, or potentially bypass, the spatial prior. Consequently, it remains unclear whether the resulting representations genuinely exploit anatomical topology or primarily benefit from increased model capacity.

The empirical comparisons rely on a relatively limited set of baselines, and several important controlled comparisons are missing. In particular, the paper does not establish the contribution of the actual electrode-distance prior against randomized or shuffled topologies, nor does it isolate the benefit of the iTransformer relative to conventional self-attention architectures.

Questions for authors

Since the proposed local graphs are full adjacency matrices over all electrodes, what mechanism ensures that they learn spatially localized or anatomically meaningful regions? How are the individual attention heads associated with the specific brain regions shown in Figure 2? Could the authors provide quantitative evidence of regional specialization?

Could the authors clarify the discrepancy between Table 1 and Table 3 regarding the average SEED-IV accuracy? Are these results obtained under identical protocols?

2 comments

Thank you for your constructive review and for recognizing the motivation of our dual-stream architecture and the complementary contributions of its components. Your comments identify useful directions for clarifying the role of spatial priors, evaluating regional specialization, and strengthening controlled comparisons. We have improved implementation reproducibility and are conducting additional SEED-IV cross-session experiments. Below, we address your concerns point by point. W1 / Q1 — Local Graphs, Regional Specialization, and Figure 2 The full N×N graph is an intentional design that preserves flexible connectivity across all electrodes. Our diversity regularization encourages different attention heads to learn distinct connectivity patterns, rather than enforcing fixed anatomical partitions. To further enhance spatial localization, we are currently exploring a spatial masking strategy that introduces explicit regional constraints into graph learning. Regarding Figure 2, the displayed patterns were selected from multiple training runs and underwent additional visualization processing. The anatomical labels were assigned descriptively based on the observed patterns, rather than through predefined regional supervision. We will clarify this distinction and further investigate quantitative measures of regional specialization. W2 — Contribution of the Spatial Prior Distance-based spatial priors have been widely adopted and validated in previous EEG graph-learning studies. Our framework builds upon this established approach while introducing learnable projections to capture task-relevant connectivity patterns. Importantly, the geometric consistency loss explicitly penalizes deviations between the learned graph and the distance-based prior, ensuring that anatomical information remains an active constraint throughout training rather than merely serving as initialization. Our ablation results further support the effectiveness of this geometric regularization. We will clarify this mechanism and its relationship to prior work in the revised manuscript. W3 — Baseline coverage and the contribution of iTransformer Our existing comparisons emphasize methods closely related to EEG connectivity modeling. We plan to broaden this evaluation with representative recent Transformer-based approaches under matched data and evaluation settings. Within our framework, iTransformer applies standard self-attention across head-specific graph representations after electrode-level GCN processing. The architectural choice concerns the representations used as tokens and the level at which their interactions are modeled. To examine this choice, we plan to compare graph-token attention with electrode-token self-attention using the same underlying GCN features and comparable encoder capacity. We will also include concatenation-based fusion without an attention encoder. These comparisons will distinguish the contribution of attention-based integration from that of organizing interactions across graph representations. We will make the token definitions and tensor dimensions explicit in the revised manuscript. Q2 — SEED-IV Averages in Tables 1 and 3 Thank you for pointing out this error. The discrepancy is due to a typographical error in Table 3. The correct average SEED-IV accuracy is 84.14%, calculated from the three cross-session results (82.23%, 84.74%, and 85.45%) reported in Table 1, rather than the incorrectly reported 85.14%. Both tables follow the same evaluation protocol. We apologize for this oversight and will correct it in the revised manuscript.

Thanks for the authors' response. I will maintain my original score.

Reviewer_QNLrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Neuro-HGLN, a neurologically informed hierarchical Graph-Transformer framework for EEG-based emotion recognition. The method jointly models global brain connectivity and region-specific local topology by combining spatial-prior-guided GCNs, parallel local graph learning, and an iTransformer for cross-region dependency modeling. Geometric and functional-diversity regularization are further introduced to preserve neurophysiological structure and encourage complementary local representations. Experiments on SEED, SEED-IV, SEED-V, and MPED show competitive cross-session and cross-subject performance with improved interpretability.

Critical strengths
  • Clear motivation and well-designed architecture. The paper explicitly models both global brain connectivity and localized topological patterns, which is well aligned with the functional segregation–integration characteristics of EEG signals.

  • Effective integration of physiological priors and data-driven learning. The spatial Euclidean prior provides anatomical constraints, while learnable global and local graphs retain flexibility for emotion-related functional connectivity modeling.

  • Comprehensive experimental validation. The method is evaluated on four EEG emotion datasets under both subject-dependent and subject-independent settings and generally achieves strong performance.

  • Reasonable ablation and visualization analyses. The ablation results support the contributions of the global/local streams and regularization terms, while learned graph visualization provides some neurophysiological interpretability.

Critical weaknesses

-The methodological novelty is somewhat incremental. The main components, including GCN, multi-head attention, iTransformer, spatial priors, and late fusion, are established techniques; the main contribution is primarily their hierarchical integration.

-The “local-region” interpretation is not fully convincing. The local graphs are learned over all electrodes rather than using explicit anatomical brain-region partitions, so describing them as region-specific neurological structures may be stronger than what the formulation directly guarantees.

-Interpretability is mainly qualitative. The paper visualizes learned graph patterns and relates them to brain regions, but lacks quantitative neurophysiological validation showing that these learned connections correspond to established emotion-related neural mechanisms.

-Generalization on the more challenging MPED dataset remains limited. Under subject-independent evaluation, accuracy is only 28.56%, suggesting that fine-grained cross-subject emotion recognition remains difficult despite the claimed robustness.

Questions for authors
  1. The local-region graphs are learned over all EEG electrodes rather than predefined anatomical regions. Could the authors clarify why these learned graphs can be interpreted as region-specific brain topology?

  2. The interpretability analysis is mainly based on visualization of learned graph structures. Can the authors provide quantitative or neuroscientific evidence showing that these connectivity patterns are consistent with known emotion-related brain mechanisms?

  3. The subject-independent performance on MPED remains relatively low. Could the authors further analyze why the model struggles on this fine-grained 7-class setting and discuss whether the limitation comes from inter-subject variability, class complexity, or dataset characteristics?

  4. Since the framework integrates several existing components, including GCNs, multi-head attention, spatial priors, and iTransformer, could the authors further clarify the main conceptual novelty beyond architectural combination?

2 comments

Thank you for your thoughtful review and for recognizing the motivation, architectural design, and integration of physiological priors with data-driven learning. We appreciate your suggestions on regional interpretation, quantitative analysis, and cross-subject generalization. We have further examined the evaluation procedures and reproducibility settings, and are conducting additional SEED-IV cross-session experiments to investigate generalization. Below, we address your questions point by point. Q1 — Interpretation of the Learned Local-Region Graphs The use of N×N graphs is deliberate rather than a lack of regional modeling. Instead of imposing fixed anatomical boundaries, our framework allows each attention head to adaptively discover connectivity patterns across the entire electrode space. The diversity loss further encourages different heads to focus on distinct connectivity patterns by reducing redundancy among their learned graphs. This enables flexible spatial specialization while preserving potential cross-region interactions. Therefore, "local-region" refers to adaptively learned regional representations rather than predefined anatomical partitions. We will clarify this distinction in the revised manuscript. Q2 — Neuroscientific Grounding and Quantitative Interpretability Previous studies have highlighted the importance of prefrontal and temporal regions in EEG-based emotion recognition, particularly in the beta and gamma bands (Zheng and Lu, 2015). Figure 2 presents selected examples from multiple training runs, with additional visualization-stage processing. The displayed patterns appear qualitatively consistent with these neuroscientific findings, suggesting that our framework can learn potentially meaningful emotion-related connectivity patterns. However, these selected visualizations do not establish consistent anatomical correspondence across runs. We will clarify the selection and visualization procedures and plan to further quantify regional connectivity distributions and cross-subject consistency in future analyses. Q3 — Subject-Independent Performance on MPED The relatively low accuracy on MPED primarily reflects the difficulty of seven-class emotion recognition under subject-independent settings. As shown in Table 2, Neuro-HGLN achieves 28.56%, compared with 28.39% for PGCN, 28.32% for TANN, and 27.95% for RGNN, indicating that this challenge is shared across representative methods rather than being specific to our framework. We attribute this difficulty to three potential factors: substantial inter-subject EEG variability, overlapping neural patterns among closely related emotional categories, and the inherent complexity of fine-grained emotion discrimination. We will further discuss these factors and the limitations of cross-subject generalization in the revised manuscript. Q4 — Main Conceptual Contribution The main conceptual contribution of Neuro-HGLN lies in the coordinated integration of geometry-guided global connectivity modeling and multiple adaptively learned local graph representations within a unified framework. Rather than simply combining existing modules, our architecture assigns them complementary roles in capturing different aspects of EEG connectivity. Specifically, the global stream models whole-scalp dependencies, while the local stream learns diverse connectivity patterns through multiple graph heads. The subsequent iTransformer enables cross-graph interaction, forming a hierarchical representation from electrode-level feature extraction to graph-level integration. Our ablation studies further support the effectiveness of these key architectural choices. We believe this coordinated design constitutes the central architectural contribution of our work. We appreciate your suggestions and will incorporate them to strengthen the interpretation, analysis, and presentation of Neuro-HGLN.

Thank you for your efforts, but I prefer to keep the original score.

Reviewer_GYBVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

EEG Emotion Recognition Using a Neural-Inspired Dual-Flow Architecture: The global stream fuses a Gaussian kernel prior map of electrode 3D coordinates with learnable matrices P and Q to form an adaptive global map, which is processed through two layers of GCN; the local stream uses multi-head attention to generate K regional-level maps from a shared latent topology. After feature extraction by K parallel GCNs, the features are concatenated and fed into an iTransformer to model cross-regional dependencies; The model is trained using a combination of two types of cross-entropy, KL geometric constraints, and inter-head diversity loss; during inference, the logits from the two streams are averaged. It achieves SOTA performance on the SEED series and MPED datasets (without requiring unlabeled test data).

Critical strengths

Ablation and visualization provide some mechanistic evidence: removing global or local flow—or both—consistently results in the disappearance of activation points; topological visualization reveals differences in activation patterns between the global map and the head-specific map, supporting the complementary hypothesis.

Critical weaknesses

W1. Lack of innovation; essentially a combination of existing components. Spatial distance prior graphs (already present in RGNN), adaptive graph learning, multi-head attention composition, parallel GCN, iTransformer (directly borrowed from time series forecasting work by Liu et al. 2023), and decision-level average fusion, each module has well-established precedents. The paper fails to demonstrate that the combination yields capabilities beyond the sum of its parts, nor does it provide a comparison with “simply concatenating these modules.” The claim of “hierarchical” is not supported by any hierarchical structure.

W2. The core neuroscientific claims are not realized by the method. The regional-level local graph is in fact a fully dense N×N graph (with no anatomical partitioning and no regional supervision), yet it claims to simulate functional separation; Figure 2 directly labels the learned heads as brain regions such as Prefrontal/L-Temporal/Occipital.This is a post hoc naming of unsupervised learning heads, constituting circular reasoning, and cannot serve as evidence of physiological interpretability.

W3. The experimental scope and baselines are weak. Only four datasets were used, all from the SEED family plus MPED; the baselines consist solely of methods from before 2023, lacking recent graph Transformer/ViT-style EEG models; MPED uses STFT features while SEED uses DE features, resulting in inconsistent inputs that make conclusions across datasets incomparable; Performance on SEED has nearly plateaued at 92–95%, leaving limited room for improvement, and the sources of error have not been analyzed.

W4. Insufficient ablation and hyperparameter tuning. Ablation was conducted on only 2 datasets with 4 variants; The number of heads (K), the number of iTransformer layers, the Gaussian kernel temperature (τ), and loss weights (γ=δ=10, α=1e-3, β=0.025—merely noted as empirically set) were not subjected to ablation; P and Q ∈ R^{N×N} cause the number of parameters to grow quadratically with the number of electrodes, yet scalability and computational cost were not discussed; no significance analysis across participants was conducted.

Questions for authors

Q1. Why is the regional-level local graph a fully dense N×N graph? What are its advantages compared to hard anatomical partitioning or direct downsampling graph construction? What is the basis for naming each head in Figure 2 after a specific brain region (are there any quantitative metrics, such as the alignment between the center of mass of activation within the head and the center of the brain region)?

Q2. What is the sensitivity of hyperparameters such as K, τ, α, and β? What are the optimal strengths for the diversity loss and KL constraint, respectively, and is there a plateau?

Q3. Why does MPED use STFT while SEED uses DE? Given that the two datasets use different features, are the ablation results (e.g., that local flow is more important for MPED) confounded by feature selection?

Q4. Why is there no comparison with EEG emotion recognition methods from 2023 onward (such as Transformer-based or channel-token ViT-based approaches)? Do the performance gains observed on existing baselines (e.g., SEED +3.6%) hold on the new baseline?

1 comment

Thank you for your detailed and constructive review. We appreciate the opportunity to clarify the architectural contributions, hierarchical design, and experimental validation of our framework. We have also strengthened reproducibility through explicit evaluation-mode control, graph recomputation upon checkpoint loading, and more complete configuration and checkpoint recording. Additional SEED-IV cross-session experiments are underway. Below, we address your concerns point by point. W1 — Architectural contribution and hierarchy (1) Architectural innovation. We respectfully disagree that combining established components necessarily implies a lack of innovation. Architectural innovation does not require inventing every individual module. Our contribution lies in the coordinated integration of geometry-guided global connectivity modeling and multiple adaptively learned local graph representations, followed by cross-graph interaction modeling. The novelty lies in this unified design rather than in any isolated component. (2) Ablation studies. The effectiveness of individual components has been extensively established in prior literature. Our ablation studies focus on validating the key architectural contributions of the proposed framework. Given the strict page limit, we believe these targeted experiments are more informative than exhaustively evaluating straightforward combinations of existing modules. (3) Hierarchical structure. We would like to clarify that "hierarchical" refers to two successive levels of representation learning: (i) electrode-level feature extraction through GCNs within individual graphs, and (ii) graph-level interaction modeling through the iTransformer across head-specific graph representations. This constitutes a hierarchy in feature representation rather than an anatomical hierarchy. W2 / Q1 — Dense graphs, regional interpretation, and Figure 2 (1) Dense graphs and diversity regularization. The N×N graph is an intentional design choice that allows each attention head to learn connectivity patterns without predefined anatomical restrictions. Unlike hard anatomical partitioning or graph downsampling, this formulation preserves all electrodes and allows flexible cross-region interactions. Importantly, our diversity regularization encourages different heads to learn distinct connectivity patterns by penalizing redundancy among their adjacency matrices. Although it does not explicitly enforce anatomical separation, it promotes differentiated spatial representations and reduces redundancy across local graphs. (2) Interpretation of Figure 2. Figure 2 presents selected results from multiple training runs with additional visualization-stage processing. The anatomical labels were assigned descriptively based on the observed spatial patterns to highlight the differentiated regional attention learned by the model. These labels were intended to facilitate interpretation rather than establish strict anatomical correspondence. We will clarify this distinction and the visualization procedure in the revised manuscript. W3 / Q3–Q4 — Feature Selection and Baseline Comparisons (1) Feature selection. The use of different features follows the established evaluation protocols of the respective datasets. DE features are widely adopted in SEED-based emotion recognition studies, whereas STFT features are commonly used in the original MPED benchmark and subsequent comparisons. We follow these conventions to ensure fair comparisons with prior methods under consistent feature representations within each dataset. Our objective is to evaluate the proposed framework across different benchmark settings rather than directly compare performance between datasets. (2) Baseline comparisons. Our baseline selection was guided by methodological relevance rather than publication year alone. Since our framework primarily addresses EEG connectivity modeling, we prioritized representative graph-learning and domain-adaptation methods, including PGCN (2023), to ensure meaningful comparisons with closely related approaches. These comparisons directly evaluate the effectiveness of our proposed architecture against established methods addressing similar challenges. Comparisons with recent Transformer-based approaches would provide additional evidence, and we plan to incorporate representative methods in the revised version. Due to space constraints, detailed hyperparameter configurations were not included in the current manuscript and will be provided in a supplementary appendix in the revised version. W4 / Q2 — Sensitivity, computational cost, and statistical evidence In Eq. (18), α and β control the geometric consistency and diversity regularization terms, respectively. Their values were determined through extensive experimental trials and comparisons rather than arbitrary empirical assignment. We plan to provide additional sensitivity analyses and experimental configurations in a supplementary appendix in the revised version.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.