EEG-Based Emotion Recognition Model Using Fuzzy Adjacency Matrix Combined with Convolutional Multi-Head Graph Attention Mechanism

Electroencephalogram (EEG) signals have been widely studied and applied in the field of human-computer interaction interfaces based on affective computing. However, the remarkable non-Euclidean structure of EEG signals poses a challenge for conventional neural network algorithms to effectively capture their complex dynamic properties. A Fuzzy adjacency matrix combined with convolutional multi-head graph attention mechanism (FCMGAT) is proposed to enhance the ability of graph neural networks to handle complex nonlinear relationships between channels during the adjacency matrix. Firstly, this method optimizes the adjacency matrix through fuzzy logic and achieves adaptive adjustment of the weight relationship between nodes through parameterized learning, thereby preventing model learning instability caused by weight oscillation; secondly, the global dependencies between channels are captured using the multi-head graph attention mechanism, which enhances the potential capturing ability of EEG emotion information. Finally, the obtained multidimensionalized emotional feature expression is input into the full connectivity layer for emotion recognition. Visual analysis of fuzzy adjacency matrices provides effective connectivity insights into affective correlations between channels. We conducted experiments on the SEED and SEED-IV datasets, and the results showed that the FCMGAT model showed significant superiority in the emotion classification of EEG signals. Its average emotion recognition accuracy reached 97.412% and 92.086%, respectively, with standard deviations of 2.385% and 3.69%. This study provides a new approach for EEG emotion recognition.