ACCNet: Adaptive cross-frequency coupling graph attention for EEG emotion recognition

• Published in: Neural networks Vol. 191; p. 107853
• Main Authors: Tian, Dongyuan, Wang, Yucheng, Gong, Peiliang, Xu, Zhewen, Chen, Zhenghua, Wei, Xiaohui, Wu, Min
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.11.2025
• Subjects: Cross-frequency coupling; Electroencephalography (EEG); Emotion recognition; Graph neural networks (GNNs); Emotion recognition; Cross-frequency coupling; Electroencephalography (EEG); Graph neural networks (GNNs)
• ISSN: 0893-6080; 1879-2782; 1879-2782
• DOI: 10.1016/j.neunet.2025.107853

EEG-based emotion recognition has emerged as a powerful approach for personalised affective computing, providing objective neural measurements for individual emotional monitoring. While Graph Neural Networks have demonstrated exceptional capability in modelling spatial relationships among EEG channels, current approaches are limited by data sparsity and uneven sampling due to fixed-frequency bands and complex cross-frequency interactions, potentially compromising the recognition accuracy and stability for single-user applications. We present ACCNet, a novel framework that enhances personalised emotion recognition through two primary innovations. First, we propose an adaptive band decomposition strategy that dynamically generates subject-specific node representations from EEG signals, enabling individualised frequency-domain analyses. Second, we introduce a cross-frequency coupling mechanism that facilitates the learning of personalised frequency relationships from a node-edge perspective, with particular emphasis on the interaction between low- and high-frequency components. The ACCNet substantially improves the capacity of graph neural networks to capture user-specific frequency interactions within EEG signals. Comprehensive empirical evaluations demonstrate that our methodology achieves superior performance in single-user emotion recognition tasks, demonstrably outperforming existing approaches. Furthermore, through specifically designed labelling noise robustness experiments, ACCNet exhibits exceptional resilience to data perturbations, validating its reliability for real-world applications. Our source code is available at https://github.com/DynamticS/ACCNet.