Representation learning of urban spatial-temporal data is fundamental and critical, serving a wide range of intelligent applications. Given that road networks and trajectories are inherently interrelated, their joint representation learning can significantly enhance the accuracy and utility of these applications. However, effectively learning joint representations for these two types of data remains challenging, particularly due to the complexities of interaction modeling and cross-scale optimization. To this end, we propose a unified framework, named UniTR, for joint representation learning of road networks and trajectories. Specifically, we first design a hierarchical propagation mechanism to model the complex many-to-many interactions between road networks and trajectories, thereby generating informative embeddings. Then, a triple-level contrastive optimization module is incorporated to systematically select valid positive and negative samples, further refining the embeddings. Experiments conducted on real-world datasets from two cities clearly demonstrate the effectiveness and superiority of UniTR.

Jie Feng, Yuwei Du, Jie Zhao, Yong Li. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Trajectory modeling, which includes research on trajectory data pattern mining and future prediction, has widespread applications in areas such as life services, urban transportation, and public administration. Numerous methods have been proposed to address specific problems within trajectory modeling. However, the heterogeneity of data and the diversity of trajectory tasks make effective and reliable trajectory modeling an important yet highly challenging endeavor, even for domain experts. In this paper, we propose TrajAgent, a agent framework powered by large language models (LLMs), designed to facilitate robust and efficient trajectory modeling through automation modeling. This framework leverages and optimizes diverse specialized models to address various trajectory modeling tasks across different datasets effectively. In TrajAgent, we first develop UniEnv, an execution environment with a unified data and model interface, to support the execution and training of various models. Building on UniEnv, we introduce an agentic workflow designed for automatic trajectory modeling across various trajectory tasks and data. Furthermore, we introduce collaborative learning schema between LLM-based agents and small speciallized models, to enhance the performance of the whole framework effectively. Extensive experiments on four tasks using four real-world datasets demonstrate the effectiveness of TrajAgent in automated trajectory modeling, achieving a performance improvement of 2.38%-69.91% over baseline methods. The codes and data can be accessed via https://github.com/tsinghua-fib-lab/TrajAgent.

Representation learning of road networks is essential for various downstream traffic-related tasks, as road network contain multi-modal data with rich information, and the learned embeddings can be directly used in machine learning models. However, due to the dynamic changes in road networks with respect to topology and associated data, as well as the local and long-range dependency caused by complex mobility semantics, learning robust and effective representations remains challenging. To this end, we exploit the properties of the road network and the mobility semantics embedded in trajectories, and propose a novel S emantic- E nhanced G raph C ontrastive L earning (SE-GCL) framework, for learning general-purpose embeddings of road networks. Specifically, in this framework, we propose (1) a multi-modal feature embedding module to capture both the attribute and visual information of road segments, (2) a semantic-enhanced graph augmentation strategy to simulate topological changes and data missing in the road network, and (3) a semantic-enhanced contrastive optimization module that leverages geo-locality and mobility semantics to guide representation learning. Extensive experiments are conducted on two real-world road networks with three representative downstream tasks. The result demonstrate that SE-GCL yields more robust and effective representations, outperforming the state-of-the-art baselines. The source code is available at https://github.com/csjiezhao/SE-GCL .

Region representations (also called embeddings) are useful for various urban computing tasks. While graph-based region representation learning methods have shown outstanding performance, they encounter two major challenges: 1) the pervasive data noise and missing data can affect the quality of the constructed region graphs; and 2) high-order relationships (i.e., group-wise relationships) among regions are often insufficiently modeled and sometimes entirely overlooked. To this end, we proposeHyperRegion, an unsupervised region representation learning framework that integrates graph and hypergraph contrastive learning to learn comprehensive region embeddings from multi-modal data. Built upon a region hybrid graph network, this framework models both pair-wise and group-wise dependencies involving POI semantics, mobility patterns, geographic neighbors, and visual semantics. To mitigate the impact of data noise and missing data, graph and hypergraph contrastive learning are performed in parallel, and a cross-module contrast is further introduced to facilitate information exchange and collaboration. Extensive experiments on real-world datasets across three downstream tasks demonstrate thatHyperRegionoutperforms all baselines, particularly improving check-in prediction by reducing MAE and RMSE by approximately 8.5% and 8.2%, respectively, and increasing$R^{2}$by about 7%.

While a variety of innovative travel modes, such as taxi service and ridesourcing service, have been launched to improve the transportation efficiency, people still encounter travel problems in real life. The major cause is the imbalance between transportation supply and demand. To strike a balance, it is well-recognized that an accurate and timely passenger demand prediction model is the foundation to enable high-level human intelligence (i.e., taxi drivers) or machine intelligence (i.e., ride-hailing platforms) to allocate resources in advance. Although quite a lot of deep models have been designed to model the complicated spatial and temporal dependencies in a data-driven way, they focus on the demand prediction of a single mode and ignore the fact that passengers may shift between different modes, especially between taxis and ridesourcing cars. In this paper, we target a co-prediction problem that considers the prediction of taxi and ridesourcing as two coupled and associated tasks, and propose a novel Temporal and Spatial Intertwined Network (TSIN) that consists of two twin components and an intertwined component. Each twin in the TSIN model is able to extract spatial and temporal dependencies from its corresponding travel mode separately (i.e., intra-mode features), and the in-between intertwined component is designed to bridge the twins and allow them to exchange information (i.e., inter-mode features), thus enabling better prediction. We first evaluate our model on four real-world datasets. Results demonstrate the outstanding performance of our model and the necessity to take into account the influence between modes. Based on an additional demand data from bike in NYC, we then discuss the generalizability in coupling more transportation modes. Further results demonstrate that our proposed intertwined neural network is highly flexible and extendable, and can yield better prediction performance.

The proliferation of multimodal big data in cities provides unprecedented opportunities for modeling and forecasting urban problems, such as crime prediction and house price prediction, through data-driven approaches. A fundamental and critical issue in modeling and forecasting urban problems lies in identifying suitable spatial analysis units, also known as city region partition. Existing works rely on subjective domain knowledge for static partitions, which is general and universal for all tasks. In fact, different tasks may need different city region partitions. To address this issue, we propose JLPR , a task-oriented framework for J oint L earning of region P artition and R epresentation. To make partitions fit tasks, JLPR integrates the region partition into the representation model training and learns region partitions using the supervision signal from the downstream task. We evaluate the framework on two prediction tasks (i.e., crime prediction and housing price prediction) in Chicago. Experiments show that JLPR consistently outperforms state-of-the-art partitioning methods in both tasks, which achieves above 25% and 70% performance improvements in terms of mean absolute error for crime prediction and house price prediction tasks, respectively. Additionally, we meticulously undertake three visualization case studies, which yield profound and illuminating findings from diverse perspectives, demonstrating the remarkable effectiveness and superiority of our approach.

Accurate traffic prediction is a critical yet challenging task in Intelligent Transportation Systems, benefiting a variety of smart services, e.g., route planning and traffic management. Although extensive efforts have been devoted to this problem, it is still not well solved due to the flexible dependency within traffic data along both spatial and temporal dimensions. In this paper, we explore the flexibility from three aspects, namely the time-varying local spatial dependency, the dynamic temporal dependency, and the global spatial dependency. Then we propose a novel Dual Graph Gated Recurrent Neural Network (DG2RNN) to effectively model all these dependencies and offer flexible (multi-step) predictions for future traffic flow. Specifically, we design a Dual Graph Convolution Module to capture the local spatial dependency from two perspectives, namely road distance and adaptive correlation. To model the dynamic temporal dependency, we firstly develop a Bidirectional Gated Recurrent Layer to capture the forward and backward sequential contexts of historical traffic flow, then combine the derived hidden states with their various contributions learned by a temporal attention mechanism. Besides, we further design a spatial attention mechanism to learn the latent global spatial dependency among all locations to facilitate the prediction. Extensive experiments on three types of real-world traffic datasets demonstrate that our model outperforms state-of-the-arts. Results also show our model has more stable performance for the flexible prediction with varying prediction horizons.