[Paper Review] Progressive Graph Convolution Network for EEG Emotion Recognition

Yijin Zhou, Fu Li+7 more|arXiv (Cornell University)|Dec 14, 2021

EEG and Brain-Computer Interfaces5 citations

TL;DR

This paper proposes a Progressive Graph Convolution Network (PGCN) for EEG emotion recognition that leverages the hierarchical nature of emotions by progressively learning from coarse- to fine-grained categories. It integrates dual graphs capturing dynamic functional connections and static spatial proximity of EEG channels, along with a dual-head module for hierarchical feature learning, achieving state-of-the-art performance on SEED-IV and MPED datasets while highlighting the role of high-frequency bands and frontal/temporal lobes in emotion processing.

ABSTRACT

Studies in the area of neuroscience have revealed the relationship between emotional patterns and brain functional regions, demonstrating that dynamic relationships between different brain regions are an essential factor affecting emotion recognition determined through electroencephalography (EEG). Moreover, in EEG emotion recognition, we can observe that clearer boundaries exist between coarse-grained emotions than those between fine-grained emotions, based on the same EEG data; this indicates the concurrence of large coarse- and small fine-grained emotion variations. Thus, the progressive classification process from coarse- to fine-grained categories may be helpful for EEG emotion recognition. Consequently, in this study, we propose a progressive graph convolution network (PGCN) for capturing this inherent characteristic in EEG emotional signals and progressively learning the discriminative EEG features. To fit different EEG patterns, we constructed a dual-graph module to characterize the intrinsic relationship between different EEG channels, containing the dynamic functional connections and static spatial proximity information of brain regions from neuroscience research. Moreover, motivated by the observation of the relationship between coarse- and fine-grained emotions, we adopt a dual-head module that enables the PGCN to progressively learn more discriminative EEG features, from coarse-grained (easy) to fine-grained categories (difficult), referring to the hierarchical characteristic of emotion. To verify the performance of our model, extensive experiments were conducted on two public datasets: SEED-IV and multi-modal physiological emotion database (MPED).

Motivation & Objective

• To address the lack of emotion-specific inductive bias in existing EEG emotion recognition models by exploiting the hierarchical structure of emotions.
• To model dynamic functional connectivity and static spatial proximity among EEG channels using a dual-graph module grounded in neuroscience.
• To improve discriminative feature learning by progressively classifying from coarse- to fine-grained emotional categories.
• To investigate the relationship between EEG frequency bands and emotional states using an interpretable deep learning framework.
• To identify brain regions most relevant to emotion expression through attention visualization and functional graph analysis.

Proposed method

• The PGCN employs a dual-graph module that combines a dynamic functional graph (learned from EEG signals) and a static spatial proximity graph (based on electrode positions) to model brain connectivity.
• A dual-head classification head is designed to progressively learn features: the first head classifies coarse-grained emotions (e.g., positive/negative), and the second refines to fine-grained categories (e.g., fear, anger, disgust).
• The model uses graph convolutional networks (GCNs) to extract spatial-temporal features from EEG signals, with attention mechanisms to emphasize relevant brain regions.
• T-SNE visualization is applied to analyze the feature space evolution during training, showing improved class separability after model optimization.
• Functional connectivity contributions are visualized via the dynamic graph $\mathbf{G_f^d}$, mapped onto the scalp to identify key electrodes and frequency bands.
• Experiments are conducted on two public datasets—SEED-IV and MPED—under both subject-dependent and subject-independent settings to validate robustness.

Experimental results

Research questions

• RQ1 Can a progressive learning strategy from coarse- to fine-grained emotional categories improve EEG-based emotion recognition performance?
• RQ2 How do dynamic functional connections and static spatial proximity among EEG channels contribute to emotion representation?
• RQ3 Which EEG frequency bands (δ, θ, α, β, γ) are most predictive of emotional states, and how do they vary across subjects?
• RQ4 Which brain regions (e.g., frontal, temporal, parietal lobes) show the highest functional connectivity contributions to emotion recognition?
• RQ5 To what extent does the PGCN model improve feature discriminability compared to baseline models?

Key findings

• The PGCN achieves state-of-the-art performance on both SEED-IV and MPED datasets, demonstrating superior classification accuracy through progressive hierarchical learning.
• The recognition accuracy for high-frequency bands (β and γ) is significantly higher than for low-frequency bands, confirming their strong correlation with emotional states.
• The δ band achieved the highest recognition rate in subject-independent experiments on SEED-IV, indicating its importance for capturing inter-subject commonalities.
• T-SNE visualization shows that after training, EEG features for the same emotion cluster more tightly, confirming improved feature discriminability.
• The frontal and temporal lobes contribute most significantly to emotion recognition, consistent with neuroscientific findings on emotion-related brain regions.
• The dynamic graph visualization reveals that functional connectivity patterns vary substantially across frequency bands, with high-frequency bands showing more distinct and localized activation.