Graph Neural Networks in EEG-based Emotion Recognition: A Survey

Chenyu Liu, Yuqiu Deng, Yihao Wu, Ruizhi Yang, Zhongruo Wang, Liangwei Zhang, Siyun Chen, Tianyi Zhang, Yang Liu, Yi Ding, Liming Zhai, Ziyu Jia, Xinliang Zhou

TL;DR

This survey establishes the first unified framework for designing Graph Neural Networks (GNNs) in EEG-based emotion recognition. By breaking down the graph construction process into node, edge, and graph-level stages, the authors organize a fragmented literature landscape and chart a course for future brain-computer interface (BCI) research.

What Problem Does This Paper Address?

Electroencephalogram (EEG) signals provide a direct, objective window into the human brain's emotional states. Because emotions are tied to coordinated neural activity across different brain regions, researchers have increasingly turned to Graph Neural Networks (GNNs) to model these complex, non-Euclidean spatial dependencies.

However, applying GNNs to EEG data isn't straightforward. Unlike standard time-series applications (like traffic forecasting), EEG graphs have unique physiological constraints—electrodes correspond to fixed physical locations on the scalp, and edges often reflect functional brain connectivity. Until now, the field lacked a cohesive taxonomy or design manual, leaving researchers to reinvent the wheel when deciding how to define nodes, calculate edges, and structure their graphs.

Methods

The authors introduce a unified three-stage framework to categorize and guide the construction of GNNs for EEG emotion recognition:

Fig. 1. Unified framework of EEG-based emotion recognition. First, (a) indicates the selection of node features. Then, (b) refers to the computation of the edge matrix representing brain connectivity patterns. Finally, (c) denotes the construction of a graph, which differs across methods.
  1. Node-level Stage: Deciding what features represent the graph's vertices. These are split into Univariate Nodes (using a single feature type like temporal domain, frequency domain, or raw signals) and Hybrid Nodes (fusing multiple feature types to capture complementary information).
  2. Edge-level Stage: Defining the relationships between electrode nodes. The paper divides these into Model-independent Edges (calculated using fixed formulas like physical distance, cosine similarity, or Pearson correlation) and Model-dependent Edges (where edge weights contain trainable parameters updated via backpropagation, such as parametric matrices or subspace projections).
  3. Graph-level Stage: Structuring the overall graph topology to capture specific neural phenomena. The authors identify four main structures:
  4. Multi-graph: Using parallel streams to capture different types of dependencies (e.g., local vs. global).
  5. Hierarchical graph: Grouping channels into functional brain regions to capture intra- and inter-region dynamics.
  6. Time series graph: Slicing the signal into temporal windows to track how the graph evolves over time.
  7. Sparse graph: Filtering out weak connections to isolate the most emotionally relevant brain activity.
Fig. 2. An overview of the categorization. The existing method was split into three stages for further categorization. These three steps correspond to (a) the selection of node features, (b) the calculation of edge matrices, and (c) the construction of graphs with different structures.

Datasets

As a survey paper, this work does not introduce a new dataset but reviews literature evaluated on standard open-source EEG emotion datasets. The most prominent benchmark discussed is SEED (SJTU Emotion EEG Dataset), which is widely noted for providing pre-computed Differential Entropy (DE) features.

Key Insights and Ideas

  • The Dominance of Differential Entropy (DE): In the node-level stage, temporal domain features—specifically DE—are the most popular and generally outperform frequency-domain features like Power Spectral Density (PSD). DE is particularly effective because it naturally measures the uncertainty and randomness of neural signals.
  • The Edge Definition Trade-off: Model-independent edges (like physical distance) provide stable, physiologically plausible constraints without adding parameters. Conversely, model-dependent edges (like subspace projections) excel at capturing complex, data-driven interactions but come with a heavy computational burden and a higher risk of overfitting.
  • Sparsity Mimics Biology: Brain activity during emotional processing is highly localized. Models that implement sparse graph structures (via thresholds or sparse weights) tend to perform better because they filter out emotionally irrelevant background noise and focus on key active regions like the frontal lobe.

Key Results

The paper provides a comprehensive mapping of existing baseline models to the proposed taxonomy. For instance:

  • Node Selection Trends: The vast majority of reviewed methods rely on univariate nodes, with raw signal inputs being the rarest due to high dimensionality and subject-specific noise.
  • Edge Selection Trends: Popular models like DGCNN and GCB-Net favor parametric, model-dependent edges to learn topologies directly from data, while models like RGNN rely strictly on physical distances.
Overview of Univariate node baselines.
Overview of Model-independent edge baselines.

Limitations and Future Work

The authors highlight four critical open challenges for the next generation of EEG-GNNs:

  1. Temporal Fully-Connected Graphs: Current time-series GNNs only link the same electrode across different time slices. They miss "hetero-electrode" relationships across time, which are vital for modeling the delayed, asynchronous responses between different brain regions.
  2. Graph Condensation: As EEG caps move toward higher resolutions (more electrodes), graphs become massive and noisy. Techniques are needed to compress these large graphs into compact, emotion-dense representations early in the training process.
  3. Heterogeneous Graphs: Human emotion involves the whole body, not just the brain. Future work should build heterogeneous graphs that connect EEG nodes with other physiological signals like heart rate (ECG) and galvanic skin response (EDA).
  4. True Dynamic Graphs: Most current "dynamic" models update edge weights via model parameters during training, but the graph structure remains static during inference. True dynamic graphs should allow nodes and edges to appear and disappear in real-time as the emotional stimulus unfolds.

Summary updated

Abstract

Compared to other modalities, EEG-based emotion recognition can intuitively respond to the emotional patterns in the human brain and, therefore, has become one of the most concerning tasks in the brain-computer interfaces field. Since dependencies within brain regions are closely related to emotion, a significant trend is to develop Graph Neural Networks (GNNs) for EEG-based emotion recognition. However, brain region dependencies in emotional EEG have physiological bases that distinguish GNNs in this field from those in other time series fields. Besides, there is neither a comprehensive review nor guidance for constructing GNNs in EEG-based emotion recognition. In the survey, our categorization reveals the commonalities and differences of existing approaches under a unified framework of graph construction. We analyze and categorize methods from three stages in the framework to provide clear guidance on constructing GNNs in EEG-based emotion recognition. In addition, we discuss several open challenges and future directions, such as Temporal full-connected graph and Graph condensation.
BibTeX
@misc{liu2024graph,
  author = {Chenyu Liu and Yuqiu Deng and Yihao Wu and Ruizhi Yang and Zhongruo Wang and Liangwei Zhang and Siyun Chen and Tianyi Zhang and Yang Liu and Yi Ding and Liming Zhai and Ziyu Jia and Xinliang Zhou},
  title = {{Graph Neural Networks in EEG-based Emotion Recognition: A Survey}},
  year = {2024},
  eprint = {2402.01138},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2402.01138},
}
Figure