Search

Traffic speed forecasting is a crucial aspect of traffic management that requires an accurate multi spatiotemporal time series forecasting technique. Previous studies typically employ graph neural network (GNN)-based methods for this task, but they are limited by their focus on spatial dependence based on real geographic distance in road networks. These structures are often inadequate for accurately describing spatial dependencies in the real world. Recently, multigraph neural networks (MGNNs) have shown considerable promise for improving forecasting performance by modelling graph structures from different spatial relationships. However, these kinds of methods do not account for complex relationships between aspects and latent dependence that cannot be known beforehand. To address these shortcomings, we propose a novel traffic speed forecasting method called adaptive multigraph convolutional networks (AMGCN), where we first introduce five predefined graphs based on spatial distance, accessibility, pattern similarity, distribution similarity and KL divergence. We fuse these graphs into a complex prior graph using a method based on spatial attention and graph relation attention. In this process, the spatial dependence in the road network is modelled comprehensively from multiple perspectives. In addition, we introduce the adaptive graph to calculate the similarity between learnable node embeddings to assist the forecasting. In this process, spatial dependencies that still cannot be captured by predefined graphs can be obtained by the way of data-driven. We utilize a mix-hop graph convolution with a residual connection to capture spatial dependencies in prior graphs and adaptive graphs. Time dependencies are also captured through causal convolution based on equidistance downsampling to prevent overfitting and redundancy in capturing spatiotemporal interactions. Extensive experiments on four real-world datasets demonstrate that our proposed method achieves superior performance compared to other baselines and effectively captures the spatiotemporal dependencies of the road network. Source codes are available at https://github.com/hfimmortal/AMGCN. [ABSTRACT FROM AUTHOR]

A smooth traffic flow is very crucial for an intelligent traffic system. Consequently, traffic forecasting is critical in achieving unwrinkled traffic flow. However, due to its spatial and temporal interdependence, traffic forecasting might be more difficult. Graph neural networks (GNN) are widely used to obtain traffic forecasting due to their capability to operate in non-Euclidean data. The topological connection of the traffic network plays a crucial role in graph structure learning. Therefore, the importance of the adjacency matrix in graph construction might be significant for effective traffic forecasting. As a result, for efficient graph construction, a Deep Spatial-Temporal Graph Neural Network (DSTGNN) is proposed for the construction of the weighted adjacency matrix and traffic speed prediction accurately by recording topological and temporal information. Three different weighted adjacency matrices are suggested depending on three different proposed algorithms for different variations of traffic's spatial condition. The new adjacency matrix is used in a novel DSTGNN for predicting the traffic state. The novel model comprised multiple layers with skip connections to adequately extract the variation of temporal and spatial information. DSTGNN outperforms the standard and other graph neural network models on four traffic speed datasets, SZ-taxi, Los-loop, PeMSD7, and METR-LA. [ABSTRACT FROM AUTHOR]

In the paper, we first explore a novel problem of training the robust Graph Neural Networks (GNNs) against noisy graphs and noisy labels. To the problem, we propose a general Self-supervised Robust Graph Neural Network framework that consists of three modules: graph structure learning, sample selection, and self-supervised learning. Specifically, we first employ a graph structure learning approach to obtain an optimal graph structure. Next, using this structure, we use a clustering algorithm to generate pseudo-labels that represent the clusters. We then design a sample selection strategy based on these pseudo-labels to select nodes with clean labels. Additionally, we introduce a self-supervised learning technique where low-level layer parameters are shared with GNNs to predict pseudo-labels. We jointly train the graph structure learning module, the GNNs model, and the self-supervised model. Finally, we conduct extensive experiments on four real-world datasets, demonstrating the superiority of our methods compared with state-of-the-art methods for semi-supervised node classification under noisy graphs and noisy labels. [ABSTRACT FROM AUTHOR]

Long-tailed methods have gained increasing attention and achieved excellent performance due to the long-tailed distribution in graphs, i.e., many small-degree tail nodes have limited structural connectivity. However, real-world graphs are inevitably noisy or incomplete due to error-prone data acquisition or perturbations, which may violate the assumption that the raw graph structure is ideal for long-tailed methods. To address this issue, we study the impact of graph perturbation on the performance of long-tailed methods, and propose a novel GNN-based framework called LTSL-GNN for graph structure learning and tail node embedding enhancement. LTSL-GNN iteratively learns the graph structure and tail node embedding enhancement parameters, allowing information-rich head nodes to optimize the graph structure through multi-metric learning and further enhancing the embeddings of the tail nodes with the learned graph structure. Experimental results on six real-world datasets demonstrate that LTSL-GNN outperforms other state-of-the-art baselines, especially when the graph structure is disturbed. [ABSTRACT FROM AUTHOR]