城市模型 (Urban Modeling)

The Scaling Bottleneck of Human Mobility Modeling

Although scaling—training larger models on larger datasets withmore compute—has proven effective for gaining performance andgeneralization across many data modalities, a comparable trendhas not yet emerged in human mobility modeling. Existing workin human mobility modeling has largely remained limited to rela-tively small models (<500M parameters) and modest data volumes(<1B samples), leaving it unclear whether, and how, scaling benefitsthis domain. This gap has limited progress toward human mobilityfoundation models for broad societal applications. In this work, weconstruct a large-scale real-world human mobility dataset compris-ing trillions of timestamped positioning records from millions ofindividuals, and take a first step toward scaling human mobilitymodeling. We first investigate whether human mobility modelsexhibit scaling behavior similar to that observed in other domains.Extensive experiments at data scales up to 1011 samples reveal anunexpected failure of conventional scaling in human mobility mod-eling. We then explore the source of this scaling failure and trace itto sample-quality issues inherent in the prevailing place-visitation(PV) modeling paradigm. Finally, we propose spatial-interactionbehavior (SIB) modeling, a new paradigm designed to overcomethis bottleneck and enable effective scaling in human mobility mod-eling. We show that SIB-based scaling recovers power-law scalingbehavior and yields models that can be directly adapted to diversedownstream tasks. These findings identify both a key obstacle toscaling mobility models and a practical route toward scalable hu-man mobility foundation models. Code and supplemental materials:https://fukamaru.github.io/human-mobility-scaling-bottleneck.

Breaking the black box: an interpretable machine learning model for global terrorism forecasting

Terrorist attacks significantly threaten a nation’s stability, prosperity, and social cohesion. Therefore, predicting terrorist attacks and identifying their underlying drivers are crucial for formulating effective counterterrorism strategies. Existing studies often prioritize either temporal or spatial dimensions, while their interplay and specific socioeconomic drivers are less explored. In this study, global news data are leveraged to construct a novel global conflict index (GCI), which integrates multisource datasets to comprehensively characterize the key drivers of terrorist attacks. TerrorXG is proposed to predict terrorist attacks, and SHAP analysis is applied to quantitatively interpret the importance and contributions of the driving factors. TerrorXG demonstrated superior performance (RMSE: 0.319; PCC: 0.777) and high computational efficiency. Compared with the second most influential factor (population size), the proposed GCI has a 42.4% greater impact on terrorist attacks. The interpretability analysis of the model highlights socioeconomic inequality as a primary determinant: the impacts of child malnutrition and infant mortality are 38.4% to 108.5% greater than the effect of urbanization. The influence of ethnicity represents only 9.7% of the impact of the GCI, providing empirical evidence that challenges traditional theoretical perspectives on ethnic conflict in terrorism research. This study provides valuable insights for optimizing the allocation of counterterrorism resources.

MGIM: Masked Geo-Inference for Land Parcels

Effective modeling of spatio-temporal contexts to support geographic reasoning is essential for advancing Geospatial Artificial Intelligence. Inspired by masked language models, this paper introduces the Masked Geographical Information Model (MGIM), a novel self-supervised framework for learning context-aware representations from multi-source spatio-temporal data. The framework’s core innovations include a parcel-scale method for multi-source data fusion and a custom self-supervised masking strategy for diverse geographic elements. This integrated modeling approach enables the model to capture complex spatio-temporal relationships and achieve consistently strong performance across diverse geographic reasoning tasks, such as trajectory inference, people flow inference, event identification, and land parcel function analysis. MGIM accurately reasons from spatio-temporal contexts and dynamically adjusts inferences according to contextual changes. The visualization of attention mechanisms further illustrates MGIM’s capacity to construct contextually-aware representations and task-specific attention patterns analogous to natural language processing models. This study presents a new paradigm for general-purpose spatio-temporal modeling in real-world geographic scenarios, offering significant theoretical and practical value, and promising an effective solution for building a geographic foundation model.

TIA-Net: Multi-Modal Land Use Recognition

With the increasing demand for refined urban management, methods that rely on a single data source or coarse-grained land use classification are no longer sufficient. Therefore, this paper proposes a parcel-level fine-grained land use recognition model, called the Triple Interaction Attention Network (TIA-Net). TIA-Net integrates remote sensing imagery (RSI), semantic information of points of interest (POI), and temporal population density (TPD). Swin-BiFPN is used to extract multi-scale spatial features from RSI. HydraMultiRocketPlus is used to model the temporal dynamics of population mobility. The POI encoder is used to characterize the distribution of human activities. Based on these components, the Feature-preserving Triple Interaction Self-Attention (FP-TISA) module is proposed. FP-TISA achieves deep fusion across spatial, semantic, and temporal dimensions. The module can effectively capture nonlinear interactions between heterogeneous data. The module can also reduce feature loss, which is common in traditional methods. On the national land use dataset CN-MSLU-100K, TIA-Net achieves a test accuracy of 77.64%, a Kappa coefficient of 0.740, and a macro-average precision of 65.20%, all outperforming the existing baseline models. Especially for macro-average accuracy, TIA-Net achieves nearly double the performance of the baseline model. Further analysis based on Grad-CAM++ and attention visualization reveals the model’s focus on key areas and its cross-modal interaction mechanism. In summary, TIA-Net improves both land use classification accuracy and interpretability. The model provides strong technical support for territorial spatial planning and natural resource management.

Multi-View Geospatial Learning for Ride-Hailing Forecasting at DiDi

The proliferation of ride-hailing services has fundamentally transformed urban mobility patterns, making accurate ride-hailing forecasting crucial for optimizing passenger experience and urban transportation efficiency. However, ride-hailing forecasting faces significant challenges due to geospatial heterogeneity and high susceptibility to external events. This paper proposes MVGR-Net (Multi-View Geospatial Representation Learning), a novel framework that addresses these challenges through a two-stage approach. In the pre-training stage, we learn comprehensive geospatial representations by integrating Points-of-Interest and temporal mobility patterns to capture regional characteristics from both semantic attribute and temporal mobility pattern views. The forecasting stage leverages these representations through a prompt-empowered framework that fine-tunes Large Language Models while incorporating external events. Extensive experiments on DiDi’s real-world datasets demonstrate the state-of-the-art performance.