Transformer-GNN Based Graph Representation Learning for Brain Functional Connectivity Network Analysis

Recently, functional magnetic resonance imaging (fMRI) has been extensively utilized to explore brain functional connectivity networks (FCNs) and leverage brain networks for predicting behavioral information. Graph Neural Networks (GNNs) have garnered significant attention in FCN analysis, due to their established power for analyzing graph-structured data. However, there are two major problems impeding the research when utilizing GNNs: (1) the limited receptive field of GNNs which only aggregate relevant information from local neighbor nodes; (2) the over-smoothing problem that arises when employing deep GNNs. To address these issues, we propose an innovative model that integrate Transformer into GNNs to enhance the ability of GNNs to aggregate relevant but relatively distant nodes. In addition, GNNs can assist Transformer to capture local neighbor node information by leveraging the graph structure information. Therefore, in this new model combining Transformer and GNNs based on the characteristics of fMRI data, we have specially designed three layers: Transformer Layer, GNN Layer and Graph Readout Layer. Finally, the experimental results on the Human Connectome Project data demonstrated the superiority of our proposed model.