To address the issues in cognitive load recognition models such as excessive reliance on manual feature extraction， ignorance of spatial information in ElectroEncephaloGram （EEG） signals， and inability to learn graph structure data effectively， a cognitive load EEG recognition model with Variational Graph AutoEncoder （VGAE） and Local-Global Graph Network （VLGGNet） was proposed. The model was consisted of two parts： a temporal learning module and a graph learning module. Firstly， the temporal learning module was employed to capture dynamic frequency representation of EEG signals using multi-scale temporal convolution， and features extracted by the multi-scale convolution were fused through Spatial and Channel reconstruction Convolution （SCConv） and 1
1 convolutional kernel cascading module. Then， the graph learning module was employed to define the EEG data as a local-global graph， where in the local graph feature extraction layer， the node attributes were aggregated into a low-dimensional vector， and in the global graph feature extraction layer， the graph structure was reconstructed using VGAE. Finally， the lightweight graph convolution operations were performed to the global graph and the node feature vector， and the prediction results were output by the fully connected layer. Through nested cross-validation， experimental results show that VLGGNet has the mean Accuracy （mAcc） and mean F1 score （mF1） improved by 4.07 and 3.86 percentage points， respectively， compared with the sub-optimal Local-Global Graph Network （LGGNet） on Mental Arithmetic Task （MAT） dataset； compared with the best-performing multi-scale Temporal-Spatial convolutional neural network （TSception） on Simultaneous Task EEG Workload （STEW） dataset， the mAcc of VLGGNet is the same as that of TSception， and VLGGNet has the mF1 only reduced by 0.01 percentage points. It can be seen that VLGGNet improves performance of cognitive load classification， and it is verified that prefrontal and frontal regions are closely related to cognitive load status.
