Fast Graph Learning for Traffic Prediction

Graph Convolutional Networks (GCNs) have been successfully utilized in modeling complex graph-structured data and have been applied in various applications such as epidemic tracing and so on. However, the training phase in GCNs faces challenges due to the computational overhead of repeated and inefficient aggregations based on graph convolution operations. We present a novel method called GCNIR that leverages reachability with incremental properties to efficiently compute diffusions in diffusion-based GCNs for node classification and traffic prediction tasks. The proposed method achieves significant speed-up gains for training semi-supervised models for node classification tasks on benchmark datasets. In addition, the proposed approach reduces the training time for diffusion-based GCN models in traffic prediction applications.