A semi-supervised co-training model for predicting passenger flow change in expanding subways

Subways in many big cities are experiencing rapid development with new lines being planned, built and put into operations. Understanding passenger flow change in expanding subways is of significant importance, because it is a crucial piece of information for the new networks’ planning and operation. Here, we employ large-scale smart card data to simulate the change of section passenger flow in an actual expanding subway. We find that the geographic properties of new subway lines can influence the passenger flows of existing sections. Based on the empirical findings, we develop a semi-supervised co-training (S-MLR-XGBoost) model to solve the problem of insufficient training samples and predict passenger flow change in the expanding subway. The proposed S-MLR-XGBoost model captures the relationship between the geographic properties of new subway lines and the passenger flow change in existing lines and achieves higher prediction accuracy comparing with several benchmark models. Finally, sensitivity analysis of the proposed model and possible reasons for passenger flow change in expanding subways are discussed.