Enhancing the interpretability of port economic modeling via implicit spatial relationship discovery

Modeling the driving mechanisms of economic activities among port cities helps reveal their interactions and spatial spillover effects, which is crucial for promoting coordinated regional economic development. Current mainstream models of these driving mechanisms are mostly based on machine learning integrated with the SHAP method, but often neglect spatial dependencies between samples—especially the implicit spatial relationships underlying port city economic activities. In recent years, AIS data has become an important tool for uncovering these implicit spatial relationships. Therefore, we propose the PortCity2Vec framework, based on AIS data and embedding representation learning, to explicitly capture implicit spatial relationships among port cities. Furthermore, we develop a spatial XGBoost model integrated with GeoShapley to incorporate these implicit spatial relationships, thereby revealing the driving mechanisms behind socioeconomic indicators and quantifying the core roles of spatial relationships and geographic features in the port economy. The results show that: (1) Implicit economic interactions among port cities extend beyond physical adjacency, indicating that port economy is influenced not only by physical proximity but also by connections within an implicit spatial structure; (2) Introducing socioeconomic indicators and geographic features of nearby port cities within the implicit spatial structure improves model accuracy, increasing R 2 from 0.6554 to 0.8541; (3) The port economy correlates positively with cargo turnover and grain crop output, but negatively with forestry and meat output. These findings highlight the key roles of economic and transport intensity and reveal resource allocation gaps. (4) Embeddings of implicit spatial relationships and geographic features effectively capture regional potential economic connections and marginal contributions that traditional models struggle to identify, thereby enhancing both the performance and interpretability of the model.