Transferability of Machine Learning Models: A Case Study of Work Trips in Eslamshahr and Qazvin

This study addresses efficient urban transportation planning through travel demand forecasting. Given the lack of comprehensive local data in many cities, model transferability is considered an important solution. Despite significant advances in machine learning, few studies have simultaneously compared the performance of linear and nonlinear models in trip generation forecasting. This research evaluates the transferability of two models, Multiple Linear Regression and Random Forest, for predicting daily household work trips, using origin-destination survey data from the cities of Eslamshahr and Qazvin, within a naïve transfer framework (without local calibration). The findings indicate that the Random Forest model achieves a coefficient of determination of 0.223 in the target dataset in the superior transfer direction (Eslamshahr to Qazvin), while this value decreases to 0.110 in the reverse direction (Qazvin to Eslamshahr). Additionally, the Multiple Linear Regression model achieves an R² of 0.245 in its best direction (Eslamshahr to Qazvin) and 0.108 in the reverse direction. The results reveal the advantage of nonlinear models under domain shift conditions and emphasize the necessity of simultaneously evaluating explanatory power and error metrics for accurate assessment of transferability. This research demonstrates the potential applicability of machine learning models in data-scarce environments.