A Novel Spatial Interaction Intensity Imputation Model Integrating Group Interaction Patterns and Distance Decay Effects

ABSTRACT Spatial Interaction Imputation aims to impute missing interaction intensity between regions based on existing interactions, thereby constructing a more comprehensive spatial interaction network for analysis. However, most existing studies focus on using neural networks to model complex pairwise (one‐to‐one) relationships between regions, generating more representative regional embeddings for interaction intensity prediction. These approaches often overlook higher‐order (many‐to‐many) interactions, such as interaction patterns between groups, which capture complex relationships at the group level and reflect large‐scale spatial interaction mechanisms. To address this limitation, we propose a Feature Fusion Encoder that integrates Graph Convolution and Hypergraph Convolution to extract both individual‐level and group‐level features, offering a more comprehensive description of regional interactions. Additionally, we incorporate distance influence into a DistMult Decoder to improve interaction intensity estimation. Validated on Beijing taxi trajectory data, our model achieves the lowest RMSE (8.06) and MAPE (26.8%) compared to baseline methods while maintaining high imputation accuracy across different distance ranges and interaction intensities. SHAP analysis further demonstrates that individual‐level features play a dominant role in spatial interaction imputation, whereas group‐level features remain indispensable. Furthermore, our study demonstrates that an appropriate hypergraph construction parameter (K) and rich regional features significantly enhance imputation accuracy. This study integrates group‐level interaction mechanisms and distance decay effects, providing novel methodologies and insights for spatial interaction network imputation and urban interaction analysis.