Synthetic Ride-Requests Generation using WGAN with Location Embeddings

Abstract

Ride-hailing services have gained tremendous importance in social life today, and the amount of resources involved have been hiking up. Ride-request data has been crucial in the research of improving ride-hailing efficiency and minimizing the cost. This work aims to model human mobility patterns to generate realistic ride-requests, addressing the prevailing problem of lack of historical training data and realistic synthetic data for different hypothetical scenarios. Synthetic generation also inherently carries anonymity. In particular, our work focuses on modeling both spatial and temporal distributions jointly for ride-hailing services. A Ride-Request Wasserstein Generative Adversarial Network (RR-WGAN) is proposed to generate plausible pick-up and drop-off geolocations. The generated ride-requests are extensively evaluated under a wide range of criteria we design, giving a comprehensive understanding of how the model performs. The proposed approach has achieved better performance than state-of-the-art methods in most scenarios. We believe this approach could provide value for ride-hailing service providers, research communities, and policy-makers.