AST-CLNet: EEG emotion recognition based on dual-branch adaptive graph convolution and contrastive learning

Electroencephalogram (EEG) emotion recognition is an important research direction in brain-computer interfaces. However, existing methods are difficult to fully capture the spatiotemporal dynamic characteristics of EEG signals and exhibit insufficient generalization ability in cross-subject scenarios with significant individual differences. To address these problems, this paper proposes a neural network model named AST-CLNet for multi-category emotion recognition, which combines dual-branch adaptive spatiotemporal graph convolution and contrastive learning. Different from traditional graph convolutional networks, the proposed model dynamically models the functional connectivity and temporal dependencies among EEG channels through learnable adaptive spatial and temporal adjacency matrices, thereby flexibly adapting to the brain topological structures of different individuals. In addition, a contrastive learning mechanism is introduced in this paper; by constructing cross-subject positive and negative sample pairs and jointly optimizing the cross-entropy classification loss and contrastive loss, the intra-class compactness and inter-class separability of feature representations are significantly improved, thus effectively enhancing the generalization ability of the model in cross-subject scenarios. The upper branch of the model extracts local spatiotemporal dynamic features via parallel spatiotemporal graph convolution, while the lower branch aggregates global context information with a channel attention mechanism, achieving effective fusion of local details and global semantics. Experiments on three public datasets (SEED, SEED-IV, and DEAP) demonstrate that AST-CLNet achieves excellent performance; specifically, on the SEED dataset, the highest intra-subject accuracy reaches 99.71% and the cross-subject accuracy reaches 87.94%, which is significantly superior to existing mainstream methods.