Key points are not available for this paper at this time.
High spatiotemporal-resolution population distribution prediction (HSTP) enhances urban management and emergency response. Existing HSTP methods commonly treat geolocated digital footprints (GDFs) as a direct population proxy, ignoring the representation bias caused by the spatiotemporal heterogeneity of geolocated behavior. Therefore, we propose HSTP, a framework that explicitly models this bias. Drawing from behavioral geography, we introduce the per capita triggering frequency of digital footprints (TFDFs), which represents the average rate at which an individual in a specific area and time generates GDF, thus capturing the local intensity of geolocated behavior. HSTP is built on the premise that GDF volume is the product of the true population and its corresponding TFDF. HSTP employs a Transformer encoder to learn geospatial context and a dual-decoder to co-predict both population and TFDF. Enforcing this premise as a training constraint allows HSTP to learn from abundant GDF data even without population labels, enhancing its spatiotemporal generalization. Experiments on hourly, 200-m-grid data from Wuhan showed that HSTP reduces prediction error (symmetric mean absolute percentage error, sMAPE) by over 35.5% compared to state-of-the-art baselines. Thus, HSTP serves as a high-precision population prediction tool and a novel framework for modeling behavioral heterogeneity in urban computing.
Xia et al. (Tue,) studied this question.