A Federated Learning-Based Framework for Ride-sourcing Traffic Demand Prediction

Abstract

Accurate short-term ride-sourcing demand prediction is vital for transportation operations, planning, and policy-making. With the models developed from data based on individual ride-sourcing companies to the joint models with data from multiple ride-sourcing companies, the prediction performance of the proposed models is enhanced significantly. However, the privacy issues of these models become a problem. Raw data collected from individual companies could cause business concerns and data privacy issues. In this paper, we propose a Federated Learning (FL) based framework for traffic demand prediction (FedTDP), to solve this problem without sacrificing the prediction performance. In our framework, the model can encapsulate the spatial and temporal correlation of traffic demand data via LSTM and GCN respectively. Moreover, by associating FL with the spatio-temporal model, no raw data is uploaded to the centralized server, and only model parameters are required. Furthermore, a Shapley value-based reward mechanism is proposed to evaluate the contribution of ride-sourcing companies and can be used as a means to distribute rewards accordingly. Finally, a real-world case study of Hangzhou City, China, is conducted. More than 16 million real-world ride-sourcing requests collected from 8 ride-sourcing companies are used, covering most of the ride-sourcing travel demand across the city. The case study shows that the FL-based spatio-temporal model outperforms several well-established prediction models while preserving data privacy. It demonstrates the effectiveness and potential of our proposed framework. Some discussions related to the real-world implementations of the Shapley value-based reward mechanism are also given in the paper.