Joint Estimation and Prediction of City-wide Delivery Demand: A Large Language Model Empowered Graph-based Learning Approach

TL;DR

This work tackles city-wide delivery demand by jointly estimating missing regional demands and forecasting future demand, including for newly planned regions and cities. It introduces IMPEL, a transferable graph-based learning framework that fuses an inductive space-then-time STGNN with LLM-derived geospatial encodings to capture region interactions and geospatial context. The method employs an end-to-end training regime with masking to enable zero-shot transfer to unseen regions and cities, achieving superior accuracy and transferability on eight cities across China and the US compared with strong baselines. By leveraging LLM-based geolocation knowledge to construct a functional graph and conditioning regional representations, the approach delivers practical benefits for scalable urban delivery planning and cross-city generalization.

Abstract

The proliferation of e-commerce and urbanization has significantly intensified delivery operations in urban areas, boosting the volume and complexity of delivery demand. Data-driven predictive methods, especially those utilizing machine learning techniques, have emerged to handle these complexities in urban delivery demand management problems. One particularly pressing issue that has yet to be sufficiently addressed is the joint estimation and prediction of city-wide delivery demand, as well as the generalization of the model to new cities. To this end, we formulate this problem as a transferable graph-based spatiotemporal learning task. First, an individual-collective message-passing neural network model is formalized to capture the interaction between demand patterns of associated regions. Second, by exploiting recent advances in large language models (LLMs), we extract general geospatial knowledge encodings from the unstructured locational data using the embedding generated by LLMs. Last, to encourage the cross-city generalization of the model, we integrate the encoding into the demand predictor in a transferable way. Comprehensive empirical evaluation results on two real-world delivery datasets, including eight cities in China and the US, demonstrate that our model significantly outperforms state-of-the-art baselines in accuracy, efficiency, and transferability.

Figures (11)

• Figure 1: This paper studies the city-wide delivery demand joint estimation and prediction problem, with the objective of estimating the demands for new regions and predicting the future demands of both existing and developing regions. We address three key challenges: (1) modeling the interaction between demand patterns of correlated regions, (2) integrating unstructured geospatial information into the demand estimator/predictor, and (3) transferring the model to new cites without re-training. The solution is developed on the foundation of an LLM-based geolocation knowledge extraction module and an LLM-enhanced spatiotemporal graph forecasting architecture.
• Figure 2: Overview of the methodology. We query LLMs using a prompt containing a basic geolocation description of the target region retrieved from OpenStreetMap. Then, the pre-activation embedding is extracted from the pre-trained LLM and fed into the downstream STGNN. The STGNN processes the vector in both node embedding and graph construction, enhancing transferability and accuracy.
• Figure 3: Overview of the solution framework.
• Figure 4: The inductive training scheme. During training stage, the observed regions are processed as subgraphs and are randomly masked for reconstruction purposes. During testing stage, the new regions are incorporated according to the prescribed graph structure and locations.
• Figure 5: Visualization examples of demand estimation (prediction) in different regions.

Theorems & Definitions (1)

• Remark