Urban spatio-temporal forecasting is critical for modern urban governance, especially in traffic management, resource planning, and emergency response. Despite advancements in pre-trained models for natural language processing, challenges persist in urban spatio-temporal forecasting. Existing methods struggle to identify and generalize universal cross-district spatio-temporal patterns, while computational limitations hinder the extraction of complex patterns from large-scale data. In this study, we propose CrossST, an efficient pre-training framework designed to capture universal spatio-temporal patterns across large-scale, cross-district scenarios. Specifically, CrossST performs pre-training on various large-scale spatio-temporal datasets to learn and store diverse valuable patterns in its pattern bank. It captures temporal dependencies, including periodicity and trends, through frequency domain and time domain analysis, while leveraging graph attention mechanisms to identify dynamic spatial propagation patterns. During fine-tuning, a spatio-temporal disentanglement strategy separates universal patterns from diverse spatio-temporal patterns stored during pre-training, improving generalization to downstream tasks and enabling efficient cross-district knowledge transfer. Additionally, temporal information aggregation and spatial linear optimization strategies enhance CrossST's efficiency and scalability, significantly reducing computational costs. Extensive experiments demonstrate that CrossST outperforms state-of-the-art baselines, improving downstream task generalization while maintaining low computational overhead. The datasets and code are available at https://github.com/Aoyu-Liu/CrossST.

Traffic forecasting plays a crucial role in intelligent transportation systems and finds application in various domains. Accurate traffic forecasting remains challenging due to the time-varying correlations within the data and the heterogeneous correlations between regions. Although various dynamic spatial-temporal graph models have been proposed to address these challenges in recent years, most of them are burdened by high computation costs and not intuitive to understand. In this paper, we propose a spatial-temporal graph model, Spatial-Temporal Dynamic Graph Diffusion Convolutional Network (SDGDN) that provides an effective and efficient approach to traffic forecasting. From the perspective of traffic flow transition probabilities, SDGDN learns dynamic graph structures to capture the time-varying traffic transition relationships. Besides dynamic graph structures, static node features are employed in diffusion convolution to better capture heterogeneous regional features. Furthermore, we utilize temporal encoding and also generate varying graphs in each stacked layer to enhance the forecasting performance. Experiments results on five real-world datasets demonstrate that SDGDN outperforms most baseline models in terms of both performance and computation efficiency.

Traffic flow forecasting is indispensable in modern urban life. Considering the complexity, variability and strong timeliness of traffic flow, traffic flow forecasting is a worth exploring but challenging research field. To achieve better traffic flow forecasting effect, we focus on two critical aspects that assume noteworthy importance: i) the features inside the traffic outflows and inflows. ii) the supplementary information regarding exterior region which is the area outside the grid division regions. To address these challenges, we propose a novel deep learning model Spatial-Temporal Flow Holistic Interaction Graph Convolution Network (STHGCN). In STHGCN, graph convolution based modules are applied through multi-step simulation. An exterior region feature estimation module is designed to estimate the influence of the special exterior region through the characteristics of complete trajectories, which enables a more comprehensive reasoning for traffic flow forecasting in grid division regions. Furthermore, a flow feature fusion integrator and stackable convolution modules are proposed to aggregate the intermediate features extracted from various perspectives, which simulate the constantly-updating and interlinked states of traffic flows through the process of multi-layer feature separation and fusion. We conduct extensive experiments on real-world traffic datasets and our proposed model outperforms all baselines.