Examining and Explaining Individual Fairness in Dynamic Pricing

Dynamic pricing has become a prevalent strategy for balancing supply and demand in urban sharing economy platforms such as Uber and Airbnb. The dynamic pricing algorithms, however, are black-boxes and have encountered issues of discrimination. While existing studies have focused on group fairness within these algorithms, limited attention has been paid to individual fairness. The key challenge is in quantifying individual similarity within temporal-spatial dimensions and the inaccessibility of the algorithms. We propose a novel framework to assess and explain individual fairness of dynamic pricing algorithms. We define individual fairness by measuring individual similarity on latent temporal-spatial representation learned from relevant downstream tasks. We also introduce a triplet loss as a fairness constraint for fair representation. As the dynamic pricing algorithms are inaccessible, we proposed a sampling based Cohort Shapley explanation method to explain the discriminatory instances. We conduct experiments on datasets from both Uber ride-sharing and Airbnb pricing platforms. Our experimental results demonstrate that our proposed triplet loss approach strikes a balance between fairness and downstream task performance. Our case study illustrates that the proposed explanation method provides reasonable and clear explanations for instances of individual unfairness.