Predicting Bus Passenger Flow and Prioritizing Influential Factors Using Multi-Source Data: Scaled Stacking Gradient Boosting Decision Trees

Accurate bus passenger flow prediction contributes to informed decisions and full utilization of transit supply. Passenger flow is affected by an extensive range of attributes featuring travel environment, which can be collected through multi-source information. A successful prediction model should not only fully utilize the latent knowledge hidden in multi-source data, but also address the resulting multicollinearity issue. Based on this principle, we propose a novel scaled stacking gradient boosting decision tree (SS-GBDT) model to predict bus passenger flow with multi-source datasets. SS-GBDT includes two modules: the prior feature-generation module and the subsequent GBDT-prediction module. The prior module entails a couple of basic models with similar performance, which generates several enhanced features of multi-source data by stacking process. Particularly, we devise a scaled stacking method by introducing a quasi-attention based mechanism (precision-based scaling and time-based scaling). Taking the newly generated features as input, the prediction module forecasts the passenger flow using GBDT model with stacked data, thereby enhancing the prediction performance. The proposed model is tested in two real-life bus lines in Guangzhou, China. Results show that SS-GBDT not only presents superiority in both prediction accuracy and stability, but can also better handle the multicollinearity issue with multi-source data. It can also prioritize the influential factors on passenger flow prediction. The prediction model is flexible and scalable, which enables the integration of various influential factors in the presence of big data.