Privacy-Preserving Short-Term Travel Demand Forecasting Based on Federated Learning

Abstract

Travel demand forecasting (TDF) is a crucial task in route planning, navigation, and scheduling. However, developing models for such tasks requires significant amount of order data, which can pose privacy concerns for users. Federated learning (FL) has addressed privacy concerns related to TDF. However, previous works only consider the data of a single factor to train the model, resulting in a less precise assessment of the demand for on-demand ride services. Moreover, non-independent and identically distributed (non-IID) data across clients also brings a challenge in personalizing models for TDF tasks when applying FL directly. In this work, a privacy-preserving short-term TDF scheme for non-IID data is proposed to deal with these problems. Our approach analyzes the correlation between various potential influencing factors and travel demand and summarizes the input variables of the FL model to achieve accurate TDF prediction with privacy preservation. Additionally, we provide each client with a personalized model trained on independent and identically distributed (IID) data to compensate for local model weight divergence. Finally, experiments are conducted on the 2022 Shanghai data of the Didi platform as a use case, and the results demonstrate that our scheme effectively improves the model's performance.