POI-Based Traffic Generation via Supervised Contrastive Learning on Reconstructed Graph

Traffic flow generation problem under realistic scenarios has raised more and more attention in recent years. This problem aims at generating traffic flow without using historical traffic data. Since road network and Point of Interest (POI) data can provide a more comprehensive picture of traffic patterns, most previous methods use both or either of them to generate traffic flow. However, roadnet graph in real-world has bias and abnormal structure, which will influence the performance of traffic generation. Previous traffic generation models directly receive real-world roadnet graph with map-match POI data as input and then use an end-to-end loss for training, which could not model the complex relationship between POI and traffic in a proper way. Different from prior methods, we propose a novel POI-based \(\underline{{\textbf {T}}}\) raffic \(\underline{{\textbf {G}}}\) eneration model via \(\underline{{\textbf {S}}}\) upervised \(\underline{{\textbf {C}}}\) ontrastive learning on \(\underline{{\textbf {R}}}\) econstructed graph, termed as TG-SCR, which combines POI data and road network data to generate the distribution of traffic flows. Our model has two novel modules: a graph reconstruction module and a POI supervised contrastive module. The graph reconstruction module includes a k-NN graph builder and a k-NN graph aggregator, which is used to reconstruct the original roadnet graph into a k-NN graph and reform POI feature. The contrastive module is used to model the relationship between POI feature and traffic flow. Extensive experiments show that our model outperforms other baseline methods on four real-world datasets.