How spatial scales enhance prediction: an interpretable multi-scale framework for bike-sharing demand prediction

Abstract

Accurate demand prediction for bike-sharing demand helps relevant authorities make informed decisions on bike placement and advance scheduling. However, most studies focus only on one specific spatial scale, thus ignoring the inter-scale synergy improvement on prediction performance. Meanwhile, the evaluation strategies through the prediction errors fail to account for the contribution of each feature to the model performance and limit the interpretability. To address these issues, we proposed an Interpretable Multi-scale framework for bike-sharing Demand Prediction (IMDP). This framework utilizes three typical scales, site scale, area scale, and global scale, to portray different but complementary knowledge on bike-sharing demands. An integrated fusion module is designed to extract and conform the spatiotemporal dependencies of each scale. Finally, a feature impact analysis strategy based on the model interpretable technique SHAP is designed to quantify and compare the feature contribution to the prediction results. Experiments on real datasets demonstrated that multi-scale information fusion improves the prediction performance and our proposed framework outperformed baselines in accuracy and interpretability. The proposed framework provides a paradigm for enhancing prediction using multi-scale spatial information and helps promote the sustainable development of urban transportation systems.