Supply-Demand Balancing and Fairness Enhancement in Ride-Sharing Services: A Graph Reinforcement Learning Based Approach

In this paper, we delve into a challenging problem within the ride-sharing service domain: how to effectively balance supply and demand while considering the dynamic behavior of both riders and drivers. Initially, we elucidate the utility functions of drivers and riders alongside their respective dynamic behavior models within the ride-sharing ecosystem. Subsequently, we introduce a Holistic Mechanism (HM) comprising a suite of algorithms grounded in Graph Reinforcement Learning (GRL). Specifically, we employ Reinforcement Learning with graph-based context to gauge the suitability of optimal pickup locations (locations which have the best chance for finding work). Furthermore, the Two-Sided Greedy Online Matching algorithm is employed to reconcile benefits between drivers and riders while also ensuring fairness among drivers. Our approach incorporates a Repositioning strategy that relocates drivers with lower utility to optimal pickup locations, with compensation provided. Fleet Size Management forces drivers to sign out when supply significantly exceeds demand and incentivizes drivers to sign in when supply is insufficient. In our experimentation, we utilize a city-scale dataset sourced from Chicago and conduct comparative analyses against various popular baseline methods. Our experimental results underscore that the GRL-HM enhances the service rate for riders and equitably dispatches drivers across the system. Furthermore, drivers can realize increased earnings, and riders experience expedited pickup times.