Using machine learning to examine freight network spatial vulnerabilities to disasters: A new take on partial dependence plots

Abstract

Analyzing transportation network vulnerabilities to disruptions is crucial for society to maintain commodity flows across the globe. However, most vulnerability analyses focus on impacts that arise from the deterioration of single network components, which can overlook spatial correlations between multiple components that manifest during area-spanning disruptions, such as those stemming from natural hazards. Here, we demonstrate an intuitive, simulation-based approach for inferring spatial vulnerabilities to area-spanning disruptions. In particular, we show how partial dependence plots derived from gradient boosting machines trained on the results of a routing simulation can be used to depict the average effect a disruption’s location has on impacts while controlling for other input variables and spatial dependencies embedded in the network. Although we demonstrate our approach for Middle Tennessee’s intermodal road and rail freight transportation network, our framework can easily be applied to other networks.