The use of the current global and regional models for predicting large-scale urban crowd flows often involves trade-offs between computational efficiency and accuracy. Global models are computationally efficient but struggle to fully capture the spatial heterogeneity in crowd dynamics and often lead to unsatisfactory performance. Region-specific models are highly accurate in capturing fine-grained spatial heterogeneity, but their computational costs are high when applied to numerous regions. We took a GeoAI approach to develop a novel spatiotemporal compressed sensing-based prediction framework (STCSP) to address these challenges. This framework employs compressed sensing techniques to identify the shared structures in crowd flow data. STCSP transforms spatiotemporal predictions in a complex geographical space into simplified predictions in an embedding space, which is more efficient than existing models. STCSP combines these simplified predictions, modeling the spatial heterogeneity in detail to increase the accuracy of crowd-flow predictions. We evaluated STCSP on a small-scale benchmark dataset and a large-scale citywide dataset and showed that STCSP outperformed 12 baseline models in accuracy and efficiency in predicting crowd flows.

Spatial heterogeneity presents significant challenges in improving the spatiotemporal prediction (STP) performance of geographical phenomena. This study identified two primary interpretations of spatial heterogeneity: variable heterogeneity, which refers to spatially uneven distributions of variables, and relational heterogeneity, which involves spatial variations in the relationships between variables. Within STP models, relational heterogeneity is a significant factor that influences performance; however, it is often overlooked by most studies that primarily focus on spatiotemporal dependency or on variable heterogeneity. The dynamics of relational heterogeneity and spatial invariance components require further investigation for predictive learning. Therefore, in this study, we developed a novel Geographical Meta-Learning Neural Network (GeoMetaNet) to address these issues. GeoMetaNet consists of a global component for spatial invariance learning and a local component for dynamic relational heterogeneity learning. In the local component, a meta-learning strategy adjusts the model parameters in different regions with various geographical environments, subject to the effects of spatiotemporal dependency and geographical similarity. We evaluated GeoMetaNet’s performance in predicting cellular traffic in Milan, demonstrating it outperformed several state-of-the-art STP models across 15 STP tasks. We analyze GeoMetaNet’s effectiveness in learning dynamic relational heterogeneity and explore geographical meta-knowledge utility for downstream mining tasks.

Forecasting the spatiotemporal distribution of mobile traffic is crucial for efficient cellular network management. Despite the superior performance of many deep learning studies, they remain inadequate for multi-city forecasting due to the neglect of geospatial effects in deep models. Specifically, spatial heterogeneity and geographical similarity suggest that distinct patterns exist within different urban regions, while shared patterns exist across different cities. To address this gap, this study proposes a Collaborative Specific-Shared Knowledge Learning (CSSKL) framework based on a meta-learning strategy for mobile traffic forecasting in two distinct cities. CSSKL consists of two key components: (1) a geographical learning module for capturing specific patterns using regional customization and (2) a geographical transfer strategy for capturing shared patterns using an attention mechanism. The effectiveness of CSSKL is validated through real-world mobile traffic datasets from two cities, namely Milan and Trentino, Italy. Experimental results demonstrate that CSSKL outperforms all baseline models, yielding a significant improvement in cross-city forecasting performance.

Geospatial interpolation plays a pivotal role in spatial analysis because it provides high-quality data support for various spatiotemporal data mining (STDM) tasks. However, statistical methods, such as kriging, face challenges in dealing with complex geo-big data. Additionally, deep-learning-based methods, despite their exceptional performance, suffer from limitations, such as poor interpretability. To harness the complementary advantages of these statistical methods and deep learning approaches, this study proposes a novel geospatial artificial intelligence (GeoAI) framework called deep kriging neural network (DKNN). The primary contribution lies in the development of an asymmetric encoder-decoder structure, which includes a deep-learning-based spatial encoder and a geostatistics-based kriging decoder. The spatial encoder consists of three specialized neural networks, whereas the kriging decoder relies on the proposed unified kriging system. During forward propagation, the kriging decoder leverage messages from the spatial encoder to generate interpolation weights for prediction. Conversely, during backward propagation, the kriging decoder guides the spatial encoder in learning interpretable knowledge. Experiments were conducted using both synthetic and practical datasets. The results demonstrate an average improvement of 20.18% in MAE, 25.04% in RMSE and 24.06% in MAPE when compared to the best-performing baseline method. Furthermore, these results confirm the superior interpretability of our DKNN framework.