Abstract

Understanding the mobility of surface freight transportation is relevant in urban planning and for developing public policies. Literature shows that most previous efforts on this topic rely on surveys and limited data. In contrast to other works, in this paper, we present an innovative methodology for characterizing last-mile freight transportation that uses a novel and copious data source: mobile phone data, which provides a broader scope. Our methodology involves calibrating supervised machine-learning models that allow us to link cell phones with truck drivers. In this endeavor, we construct several input variables that track mobile phone's daily movement patterns, including traveled distances, interactions with highway networks, and land use variables. We test our approach by conducting a case study in Santiago, Chile, for which we analyze mobility patterns and logistics indicators disaggregated by day, hour, and zoning. For this case, we show that all supervised models performed well regarding AUC, which can be attributed to the high granularity and handling of the data. However, we chose to use NGBoost in all subsequent experiments, as it provided slightly better results on our validation data. Our work has several implications for practice. For instance, our results can support decision-makers and policymakers in identifying critical areas where urban logistics centers and transportation interventions are needed. Finally, several research lines stem from our work, which include assessing the impact of incorporating geospatial information and the measurement of logistics sprawl over time.