MTUKG: A Multi-scale Temporal Urban Knowledge Graph Dataset for Knowledge-Enhanced Spatiotemporal Prediction
TL;DR
Experiments on USTP and knowledge reasoning tasks demonstrate that MTUKG significantly enhances predictive performance and provides a structured knowledge resource for modeling the temporal dynamics, functional evolution, and multi-scale dependencies in urban systems.
Abstract
Integrating multi-source knowledge into urban knowledge graphs (UrbanKGs) has shown strong potential for improving urban spatiotemporal prediction (USTP). However, existing public UrbanKGs are often misaligned with the temporal and multi-scale nature of USTP. First, they usually represent urban facts as static triples, ignoring when facts become valid, expire, or change over time. Second, their functional-zone layers are typically derived from static POI clustering, which overlooks spatial boundaries, activity interactions, and event-driven temporal semantics. As a result, current datasets provide limited support for modeling temporally valid knowledge and its cross-scale effects on prediction. To address these limitations, we propose MTUKG, a multi-scale temporal urban knowledge graph dataset. MTUKG represents urban events as time-interval facts, enabling knowledge to be aligned with specific prediction windows. It further introduces a spatial unit-driven method that integrates spatial morphology, semantic distributions, spatial decay effects, and dynamic events to construct temporal functional zones and establish multi-scale temporal associations across micro, meso, and macro urban scales. Experiments on USTP and knowledge reasoning tasks demonstrate that MTUKG significantly enhances predictive performance and provides a structured knowledge resource for modeling the temporal dynamics, functional evolution, and multi-scale dependencies in urban systems. The dataset and source code are available at https://github.com/Jiahaohuhu/MTUKG.