Abstract:Accurate spatiotemporal forecasting is crucial for optimizing traffic resource allocation and air pollution warnings. Recent studies describe the spatial correlations between stations with the help of pre-defined graphs and then use graph convolutional network (GCN) to learn node representations by aggregating neighborhood information of all nodes uniformly to capture spatial dependencies. However, pre-defined graphs can not accurately describe the complex spatial correlations between stations. In addition, these GCN-based models not only suffer from over-smoothing issues when stacking multiple GCN layers but also exhibit limited structure-aware (i.e., fewer links) on low-degree nodes, leading to nodes capturing contextual information on local spatial proximity. To this end, in this paper, we propose a novel learning-based spatial graph, which can effectively learn the spatial collaborative correlations between stations, including geographic proximity and functional similarity (e.g., two stations both located in residential areas, thus sharing strong regional dependencies), and represent them as an adjacency matrix. Moreover, we propose a fine-grained graph convolutional network, named FGGCN, aimed to enhance the representations of low-degree nodes in the spatial graph. The core advantage is that FGGCN aggregates neighborhood information for the differences in structure-aware of high- and low-degree nodes, respectively, and adaptively learns the importance weights of different layer representations, allowing the nodes to aggregate the abundance global contextual information. We further incorporate the spatial representations captured by FFGCN into an Encoder-Decoder architecture, recursively capturing spatiotemporal autocorrelation and generating multi-step predictions. We extensive experiments on five real-world spatiotemporal datasets, covering air quality and bike flow. The experimental results demonstrate the state-of-the-art of our propose method.

Graph Learning for Spatiotemporal Signals with Long- and Short-Term Characterization

Graph Learning Based on Spatiotemporal Smoothness for Time-Varying Graph Signal

Graph Learning for Spatiotemporal Dynamic Signal

Spatiotemporal Smoothness-based Graph Learning Method for Sensor Networks

INCREASE: Inductive Graph Representation Learning for Spatio-Temporal Kriging

Graph-Free Learning in Graph-Structured Data: A More Efficient and Accurate Spatiotemporal Learning Perspective

Graph signal reconstruction based on spatio-temporal features learning

Mining Spatio-Temporal Relations Via Self-Paced Graph Contrastive Learning

Learning Time-aware Graph Structures for Spatially Correlated Time Series Forecasting

Learning Spatiotemporal Graphical Models From Incomplete Observations

Graph Laplacian Matrix Learning from Smooth Time-Vertex Signal

Decoupling Long-and Short-Term Patterns in Spatiotemporal Inference

Time-varying graph learning from smooth and stationary graph signals with hidden nodes

Spatio-Temporal Signal Recovery Based on Low Rank and Differential Smoothness

Fine-grained Graph Convolutional Network with Learning-Based Bi-Relational Graph for Spatiotemporal Forecasting

Sparse Graph Learning from Spatiotemporal Time Series

Dynamic Graph Structure Learning for Multivariate Time Series Forecasting

Correlation-aware Spatial-Temporal Graph Learning for Multivariate Time-series Anomaly Detection

Learning Decomposed Spatial Relations for Multi-Variate Time-Series Modeling.

Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective

Time-Varying Graph Signal Recovery Using High-Order Smoothness and Adaptive Low-rankness