LGSAN: Latent Graph Structure Aware Network for Traffic Forecasting

Abstract:
Precise traffic prediction, the cornerstone of intelligent traffic systems, has never been more crucial than it is now, thanks to the surge in smart cities and urban computing. Traditional approaches have recently been surpassed by graph neural networks. Most classic GNN-based models, on the other hand, perform well when given a preset graph topology. Existing approaches for defining network structure only consider spatial relationships and disregard temporal dependencies. Moreover, the semantics of static preset graph adjacencies used during training are always inadequate, neglecting the fundamental topology that may help the model be fine-tuned. To overcome these issues, we present the Latent Graph Structure Aware Network (LGSAN), a new traffic prediction system. The model uses a multi-layer perceptron and a KNearest Neighbor-based graph generator to acquire latent network topology information from the full data set while taking into account both spatial and temporal dynamics. Furthermore, by initializing MLP-kNN-hybrid based on true adjacency matrix and similarity measure by kNN, LGSAN aggregates topologies concentrating on geographic and node similarity. The produced graph is also fed into the spatiotemporal basic feasible, which combines diffusion graph convolution with a gated temporal convolutional network. LGSAN outperforms several kinds of state-of-the-art baselines in real-world experiments using two benchmark datasets.