Fo Hu ¹ Kailun He ¹ Mengyuan Qian ¹ Xiaofeng Liu ¹ Zukang Qiao ² Lekai Zhang ^3* Junlong Xiong ^2*

• ¹ College of Information Engineering, Zhejiang University of Technology, Hangzhou, Jiangsu Province, China
• ² First Affiliated Hospital, Zhejiang Chinese Medical University, Hangzhou, China
• ³ Zhejiang University of Technology, Hangzhou, Zhejiang Province, China

Emotion recognition based on electroencephalography (EEG) is a pivotal component of braincomputer interface research. To attain precision in emotion recognition from EEG signals, it's crucial to adeptly extract and integrate both their spatial and temporal attributes. Nevertheless, prevailing research often concentrates on one aspect, neglecting the integration of spatial and temporal dynamics fully. To address this limitation, we propose a novel Spatiotemporal Adaptive Fusion Network (STAFNet) that enhances both the accuracy and robustness of EEGbased emotion recognition. The model incorporates an adaptive graph convolutional module, which leverages an adaptive adjacency matrix to thoroughly capture brain connectivity patterns through spatial dynamic evolution. In addition, we introduce a multi-structured transformer fusion module that fully exploits the latent correlations between spatial and temporal features, yielding a fused representation for emotion classification. Extensive experiments were conducted on two public datasets, SEED and SEED-IV, to evaluate the model's performance. The results demonstrate that STAFNet achieved accuracies of 97.89% and 93.64% on SEED and SEED-IV, respectively, outperforming current state-of-the-art methods. Our findings highlight the critical role of spatiotemporal feature extraction in emotion recognition and introduce a novel framework for feature fusion.