A spatial–temporal graph-based AI model for truck loan default prediction using large-scale GPS trajectory data

Abstract

With the increasing uncertainties in freight transportation, truck loans are playing a crucial role in the stability and development of the logistics industry. A pivotal problem to truck loan management is controlling credit risk. Much research has focused on pre-loan default prediction considering applicants’ static personal information. However, few studies concentrate on post-loan risk control, which needs to track customers’ dynamic behaviors to ensure the loans can be recovered. We propose a Spatial–temporal Graph-based AI model for Truck loan Default prediction (SGTD) in post-loan management, which considers spatial and temporal dependencies of customers’ dynamic behaviors using large-scale GPS trajectory data. (1) The spatial dependency is caused by the functional similarity of the trajectories, which is only relevant to people’s travel habits and purposes. Therefore, we detect stay points from the nationwide trajectories and represent them with point-of-interest (POI) information as POI nodes. Then, we learn the latent vectors of POI nodes in a heterogeneous network through metapath learning, which are utilized to initialize vectors in modeling the temporal dependencies. (2) To learn temporal dependencies, we split the long trajectory into period sequences according to the cyclical nature of truck loans. Then we model each sequence with graph neural networks to capture the short-term closeness and period dependencies. Finally, we introduce LSTM networks to learn the long-term trend dependency and predict the default probability. The outcomes of extensive experiments on a real-world dataset demonstrate the effectiveness of SGTD in AUC, KS, accuracy, and F1-score, as well as the practical economic contribution. The results also implicate that GPS data can contribute to learning drivers’ socioeconomic information, and the proposed method is applicable to other transportation scenarios.