Adaptive Progressive Attention Graph Neural Network for EEG Emotion Recognition

Tianzhi Feng, Chennan Wu, Yi Niu, Fu Li *, Yang Li *, Boxun Fu, Zhifu Zhao, Xiaotian Wang · 2025

Read Adaptive Progressive Attention Graph Neural Network for EEG Emotion Recognition by Tianzhi Feng, Chennan Wu, Yi Niu, Fu Li *, Yang Li *, Boxun Fu, Zhifu Zhao, Xiaotian Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-14246.txt.

Introduction

Tianzhi Feng and co-authors present in 2025 a graph neural network that reads emotion from brain wave signals. Brain regions linked to emotion form shifting networks, and those networks differ across people and moods. A model that treats the regions as a fixed map will miss that motion, so theirs rebuilds the map as it goes.

Three expert modules work in steps: one reads the whole brain pattern, one zooms on single regions, and one picks out the channels most tied to emotion. A weight maker then blends the three views into the final label. Tests on three public emotion data sets, covering varied numbers of emotion classes, put the method ahead of baseline rivals. The layered design mirrors how emotion builds from broad state to sharp focus, and it adapts to each head it reads.

Catalog record

CREATORTianzhi Feng, Chennan Wu, Yi Niu, Fu Li *, Yang Li *, Boxun Fu, Zhifu Zhao, Xiaotian Wang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-14246

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Adaptive Progressive Attention Graph Neural Network for EEG Emotion Recognition Tianzhi Feng, Chennan Wu, Yi Niu, Fu Li *, Yang Li *, Boxun Fu, Zhifu Zhao, Xiaotian Wang 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. I Introduction Emotions significantly impact various aspects of human daily life and psychological health, influencing our decision-making, motivation, attention, memory, problem-solving abilities [ 1 , 2 ] . Consequently, it is necessary to develop methods to objectively and accurately identify human emotions. The ability for machines to understand human feelings has become an important field of study, drawing significant attention from Human–Machine Interaction (HMI) and pattern recognition researchers lately [ 3 , 4 , 5 , 6 ] . Most of these studies primarily utilize two typical signals, i.e., external and internal responses. External responses mainly include some behavioral data, such as facial expression [ 7 ] , speech signals [ 8 ] , conversational data on social media platforms [ 9 ] . Internal responses are based on physiological signals, including electromyography (EMG) [ 10 ] , electrocardiogram (ECG) [ 11 ] , and electroencephalogram (EEG) [ 12 ] . Neuroscience research suggests that physiological signals provide more direct access to emotional origins than behavioral indicators. Thus, more and more researchers focus on this field during the past several years. EEG is a widely used technique for recording the electrophysiological activity of neurons in the cerebral cortex via electrodes attached to the scalp [ 13 ] . As a typical physiological signal, EEG has shown considerable promise in decoding human emotions [ 14 , 15 , 16 , 17 ] . Over the past decades, numerous approaches have emerged for efficiently decoding emotions from EEG signals. Among various methods, convolutional neural networks (CNNs) are extensively employed in recognizing emotions from EEG data. For example, Hasan et al. [ 18 ] utilize Fast Fourier Transform (FFT) and CNNs to classify 64 emotions, achieving notable accuracy in valence and arousal dimensions. Similarly, Ali et al. [ 19 ] utilize a Capsule Network (CapsNet) that outperforms traditional support vector machines (SVMs) and CNNs, achieving average accuracies of 80.22% and 85.41%, respectively. Ahmad et al. [ 20 ] focused on a CNN-based model for classifying emotions into positive, neutral, and negative categories, attaining over 93% accuracy across all categories. Recent advancements include He et al.’s [ 21 ] CNN architecture for complex EEG signal classification and Aldawsari et al.’s [ 22 ] optimized 1D-CNN-based process, which significantly improve efficiency. Hybrid models combining CNNs with other techniques, such as LSTM and ensemble learning, have also shown promising results. As an example, Ali et al. [ 23 ] and Yuvaraj et al. [ 24 ] demonstrate high accuracies, with Ali’s model reaching an impressive 98%. Other notable innovations include the integration of differential entropy with CNN-BiLSTM by Cui et al. [ 25 ] , which achieves over 94% accuracy on DEAP and SEED datasets, and the introduction of CIT-EmotionNet by Lu et al. [ 26 ] , which combines CNN and Transformer models to outperform previous methods. Huang et al. [ 27 ] incorporate attention mechanisms within a CNN-BiLSTM framework, achieving near-perfect accuracies in multi-class tasks. Further studies, such as those by Saha et al. [ 28 ] , Li et al. [ 29 ] , and Tao et al. [ 30 ] , explore wavelet decomposition, multi-scale CNNs, and attention mechanisms, all contributing to improved emotion recognition performance from EEG signals. Wang et al. [ 3 ] introduces self-supervised learning to CNNs, enhancing both resource utilization and performance. Recent approaches include Jin et al.’s [ 31 ] CNN-Transformer network for fNIRS-based emotion recognition and Farokhah et al.’s [ 32 ] simplified 2D CNN with selective channel to improve inter-subject accuracy. Asif et al. [ 33 ] demonstrate the potential of combining various convolution layers within a CNN framework for subject-independent tasks. While these algorithms have shown impressive performance, they often overlook the inter-channel topological relationships inherent in EEG channel spatial features. This limitation arises from CNNs’ inability to process non-Euclidean spaces, such as graphs and manifolds. To better capture the spatial relationships within EEG signals, several Graph Neural Network (GNN)-based approaches have been proposed, yielding promising results. Zheng et al. [ 34 ] propose a Hierarchy Graph Convolution Network (ERHGCN) that achieves classification performance of 90.56% for valence dimension and 88.79% for arousal dimension, demonstrating the potential of hierarchical GCN structures for emotion recognition. Saboksayr et al. [ 35 ] apply Graph Signal Processing (GSP) techniques, showing enhanced performance when compared with traditional methods and highlighting the efficacy of graph-based approaches. Gilakjani et al. [ 36 ] combine GNNs with contrastive learning and GAN-based data augmentation, leading to enhanced classification performance on both DEAP and MAHNOB-HCI datasets. Li et al. [ 29 ] develop a model that learns discriminative graph topologies in EEG networks, achieving an average accuracy of 84.56% in online experiments, which is particularly important for affective brain-computer interfaces. Klepl et al. [ 37 ] present a comprehensive survey on GNN applications in EEG signal classification, providing valuable insights into their advantages and limitations. In the realm of deep learning, Abdulrahman et al. [ 38 ] achieve notable accuracies of 70.89% in binary classification and 90.33% for multi-class emotion recognition tasks using advanced deep learning models. DGCNN, proposed by Song et al. [ 39 ] , dynamically learns the graph’s adjacency matrix to establish spatial relationships. Zhong et al.’s RGNN [ 13 ] enhances GNN model robustness against cross-subject variations and noisy labels by incorporating two regularization techniques. The IAG model by Song et al. [ 40 ] adaptively generates directed graph connections from input graphs, allowing for exploration of intrinsic relationships between EEG regions. Although considerable progress has been made in emotion recognition using EEG, current approaches still face challenges in dynamically capturing the complex relationships between emotional patterns and brain functional regions. One major challenge is the significant variation in emotion-related brain activity between individuals, which demands more adaptive recognition approaches. To address these challenges, we propose a novel Adaptive Progressive Attention Graph Neural Network (APAGNN) that finely screens out critical EEG channels across different subjects through three specialized experts. Specifically, three experts work collaboratively through a progressive analysis pipeline. The first expert analyzes global brain topology patterns, followed by the second expert that identifies emotion-specific regions, while the third expert focuses on critical channels. This hierarchical refinement process

Read the complete text