Abstract: Understanding freight mode choice preferences is vital for optimizing supply chain efficiency and guiding strategic infrastructure investments. Existing models primarily focused on shipment-specific data such as commodity weight and value, yet the logistical factors such as the ease of first mile and last mile to terminals are often overlooked. These factors implicitly impact good transfer efficiency, operational costs, and service reliability thereby influencing the shippers' choices. This study introduces a novel framework that combines localized geographic insights with ensemble learning techniques. Specifically, we first quantified the accessibility of facilities associated with different commodity groups to key terminals, such as trucking terminals and cargo airports. These accessibility measures were then used to derive origin and destination spatial embeddings via Graph Convolutional Networks (GCNs) thereby capturing the spatial dynamics of freight movement and the underlying shipment origin-destination patterns. Thereafter, these GCN-derived features complement shipment data and feed into an ensemble model that leverages predictive strategies through both voting and stacking mechanisms. Two types of models were developed to validate the importance of incorporating spatial features: the benchmark model that relies exclusively on conventional shipment metrics and our proposed model that integrates spatial features via GCN-derived embeddings. The results indicate that our model incorporating spatial embeddings significantly outperforms the model based solely on shipment-specific data, practically for low frequency modes such as “Air”. The proposed model achieves up to a 27% improvement in prediction accuracy, which demonstrates the importance of involving spatial dynamics of freight transportation. This refined approach improves freight mode choice predictions with key insights. Our study could be used to inform strategic investments in transportation infrastructure and results in more efficient and responsive freight systems.

Mohammad Khojastehpour is a Ph.D. student in Transportation Engineering at the University of Tennessee, Knoxville. His research focuses on traffic operations, freight transportation, and transportation system resilience, with an emphasis on data-driven analysis of real-world problems. He has worked on projects involving traffic monitoring data, freight network performance, roadway incidents, and infrastructure disruptions.

His recent work includes studying freight network resilience during major bridge closures, analyzing the operational impacts of incidents on freeway systems, and contributing to shipper behavior studies. His interests include freight transportation, network resilience, and improving transportation system performance through practical analytical methods.