Presentation Information
[1Yin-B-17]QSTMF: Censored Demand Prediction via Survival Analysis with Locally Stationary Poisson Arrivals and Low-Rank Structure
〇Hideki Hayashi1, Kazushi Ikeda1 (1. NARA Institute of Science and Technology)
Keywords:
Censored demand estimation,Queueing theory,low-rank factorization,Shared mobility,Survival analysis
Accurate latent-demand forecasting is essential for rebalancing and capacity planning in shared-mobility services such as bike sharing. However, observed rental counts are censored by inventory constraints, and estimators based only on counts systematically underestimate demand. This paper assumes a locally stationary Poisson arrival process and uses event times at which waiting vehicles are rented out within each time slot. We model the event-time distribution and treat unused initial inventory as right-censoring in a survival-analysis framework, then estimate demand intensity by maximum likelihood. We further impose a low-rank structure on the intensity matrix to complete station–time slots with zero inventory and no observations. Simulation results show that the proposed method reconstructs latent demand more accurately than existing approaches under severe censoring.