Fairness in Online Jobs

Abstract

Online job marketplaces are becoming very popular. Either jobs or people are ranked by algorithms. For example, Google and Facebook job search return a ranked list of jobs given a search query. TaskRabbit and Fiverr, on the other hand, produce rankings of workers for a given query. Qapa, an online marketplace, can be used to rank both workers and jobs. In this thesis, we develop a unified framework for fairness to study ranking workers and jobs. We case study two particular sites: Google job search and TaskRabbit. Our framework addresses group fairness where groups are obtained with any combination of protected attributes. We define a measure for unfairness for a given group, query and location. We also define two generic fairness problems that we address in our framework: quantification, such as finding the k groups (resp., queries, locations) for which the site is most or least unfair, and comparison, such as finding the locations at which fairness between two groups differs from all locations, or finding the queries for which fairness at two locations differ from all queries. Since the number of groups, queries and locations can be arbitrarily large, we adapt Fagin top-k algorithms to address our fairness problems. To evaluate our framework, we run extensive experiments on two datasets crawled from TaskRabbit and Google job search.