Evaluating Algorithm Fairness in an Agent-Based Taxi Simulation

The popularity of ride-sharing services, such as Uber and Lyft, is quickly transforming urban transportation systems. With the digitalization of such services and taxi dispatching systems, companies are now able to optimize their matching algorithms to maximize profit while reaching the best possible user satisfaction in terms of service availability or waiting time. Unfortunately, little attention is being paid to possible negative consequences incorporated into these matching systems from the drivers’ perspective. Recent studies raise concerns about risks threatening workers’ well-being, including racial bias, worker safety, fairness to workers, and asymmetries of information and power. While most literature in the area of taxi matching algorithms is concerned with optimizing aggregate outcomes for the system, we focus on a so far neglected angle, namely fairness from the drivers’ perspective]. Similarly to recent work in the algorithmic fairness community, our goal is to guarantee an even distribution of benefits to the workers on these platforms.
We use an agent-based model system over a regular city grid, where taxi drivers and users are paired using different algorithms with several choices of system parameters. We investigate the overall sum and the distribution of total taxi earnings over time along with multiple other metrics, and we introduce a numerical measure for the fairness of the income distributions. First, we explore how the fairness changes using different simplified city geometries having one or multiple ‘attractive centers’. We also change the overall density of taxis and users present in the city. The results suggest that the geometry, user/taxi ratio, or the change of density alone can lead to growing inequalities, thus decreasing the fairness of the system. Second, we design matching algorithms with additional fairness constraints based on various matching algorithms borrowed from real platforms or the taxi literature. We then explore the tradeoff between fairness to workers, and the benefits to customers or the company. Our simulations suggest that it is possible to decrease the inequality among drivers while keeping the high overall income of existing solutions. Incorporating fairness as an optimization criterion could thus provide a framework for external monitoring and evaluation of matching algorithms in ridesharing services.