Predicting bike-sharing demand using light gradient boosting machine and Shapley additive explanations values: a cross-regional generalisability study

The promotion of sustainable travel methods, such as public transportation, walking and bike-sharing, is being carried out in many countries around the world to raise awareness of the harmful effects of motorised traffic on the environment and form sustainable travel habits. Bike-sharing is considered a valuable option as it contributes to emission-reduction goals. This study investigates the transferability of a unified light gradient boosting machine (LightGBM) framework for bike-sharing demand prediction across three distinct socio-economic and climatic urban archetypes, namely Seoul, Washington D.C. and London, using variables including temperature, humidity, wind speed, season, hour, working day or holiday and location. While previous research focuses on localised models, this study tests the hypothesis that a single, high-fidelity model can transcend geographical heterogeneity. The results, validated through ten-fold cross-validation to ensure robustness, show that the predictive LightGBM model has a coefficient of determination, R2, of 0.947, root mean square error of 195.532 and mean absolute error of 107.548. Shapley additive explanations interpretability reveals that while temporal cycles and thermal comfort are universal predictors, the location feature captures latent socio-technical maturity, where London exhibits significantly higher peak-hour demand intensity compared to Washington D.C. and Seoul.