KAN-GCNN: EEG-Based Emotion Recognition with a Kolmogorov-Arnold Network-Enhanced Graph Convolutional Neural Network

Emotion recognition, relying on electroencephalogram (EEG), has evolved into a critical technology within the domain of affective computing. However, many existing studies have not fully exploited the topological structure between EEG electrodes, and the recognition methods are often confined to traditional machine learning approaches or multi-layer perceptron (MLP)-based neural networks. In contrast to MLP, the Kolmogorov-Arnold Network (KAN), which is built upon learnable activation functions, has the potential to offer superior performance in both accuracy and interpretability. Therefore, this study proposes a Kolmogorov-Arnold network-enhanced graph convolutional neural network (KAN-GCNN) model for EEG-based emotion recognition. The KAN-GCNN utilizes a graph structure to represent multi-channel EEG signals, employing a graph neural network (GNN) based on KAN for emotion recognition. This model incorporates learnable parameters for both the 'edges' of the GNN and the KAN, which enables more precise capture of local and global interactions between EEG channels. Comprehensive experiments were performed on the SEED (three-class emotion) and the SEED-IV (four-class emotion). The experimental results reveal that the KAN-GCNN outperforms the current leading techniques in emotion recognition accuracy. In subject-specific experiments, the KAN-GCNN attained average recognition accuracies of 97.64% and 85.65% on the SEED and SEED-IV, respectively. These outcomes further demonstrate that the learnable activation function in KAN significantly enhances the model's flexibility and adaptability, leading to improved emotion recognition accuracy.