Adaptive Progressive Attention Graph Neural Network for EEG Emotion Recognition

TL;DR

The paper addresses EEG emotion recognition under substantial inter-subject variability by introducing APAGNN, a three-expert progressive attention graph neural network that models brain topology from global to region to electrode scales. It incorporates a diversity-preserving Jensen-Shannon objective and a dynamic fusion mechanism to adaptively weigh expert outputs, enabling personalized and robust predictions. Across SEED, SEED-IV, and MPED, APAGNN achieves state-of-the-art accuracies with reduced variance, and extensive attention and t-SNE analyses provide interpretable links between learned channels and brain regions. The work advances EEG-based affective decoding by coupling adaptive, multi-scale graph reasoning with explainable channel importance, offering practical benefits for HMI and brain-computer interfaces.

Abstract

In recent years, numerous neuroscientific studies demonstrate that specific areas of the brain are connected to human emotional responses, with these regions exhibiting variability across individuals and emotional states. To fully leverage these neural patterns, we propose an Adaptive Progressive Attention Graph Neural Network (APAGNN), which dynamically captures the spatial relationships among brain regions during emotional processing. The APAGNN employs three specialized experts that progressively analyze brain topology. The first expert captures global brain patterns, the second focuses on region-specific features, and the third examines emotion-related channels. This hierarchical approach enables increasingly refined analysis of neural activity. Additionally, a weight generator integrates the outputs of all three experts, balancing their contributions to produce the final predictive label. Extensive experiments conducted on SEED, SEED-IV and MPED datasets indicate that our method enhances EEG emotion recognition performance, achieving superior results compared to baseline methods.

Figures (5)

• Figure 1: The architecture of APAGNN. In the multi-expert architecture, each expert learns discriminative features for emotion classification while generating attention maps to guide subsequent experts. The first two experts progressively refine the graph structure through attention operations and node pruning, optimizing the input for the next expert in the sequence. The dynamic expert fusion module then integrates the experts' predictions using a weight generator to produce the final model output.
• Figure 2: Schematic illustration of the expert module. Each expert performs two parallel tasks: (a) graph convolution operations for feature learning and (b) attention map generation for channel importance weighting. The module produces dual outputs: classification decisions and corresponding attention maps that highlight the influential EEG channels contributing to these decisions.
• Figure 3: Confusion matrices for subject-dependent experiments on SEED, SEED-IV and MPED datasets.
• Figure 4: Attention maps generated by three experts across different subjects from three datasets: (a) SEED, (b) SEED-IV, and (c) MPED. Each row represents a different subject, while columns show the attention patterns generated by each expert in sequence. For completeness of analysis, we visualize patterns from three experts, although only the first two are implemented in the APAGNN.
• Figure 5: T-SNE visualization of the feature learning process on the MPED dataset. Each row shows the feature evolution for an individual subject across four stages: initial normalized EEG data (first column) and subsequent transformations by each expert (columns 2-4). Seven emotions are represented by distinct colors: joy (red), funny (blue), neutral (yellow), sad (pink), fear (purple), disgust (black), and angry (green).