Quantifying the Impact of User Attention on Fair Group Representation in Ranked Lists
share
In this work we introduce a novel metric for verifying group fairness in ranked lists. Our approach relies on measuring the amount of attention given to members of a protected group and comparing it to that group's representation in the investigated population. It offers two major developments compared to the state of the art. First, rather than assuming a logarithmic loss in importance as a function of the rank, we allow for attention distributions that are specific to the audited service and the habits of its users. For example, we expect a user to see more items during a single viewing of a social media feed than when they inspect the list of results for a single query on a web search engine. Second, we allow non-binary protected attributes to enable investigating inherently continuous attributes (for example political alignment on the Democratic vs. Republican spectrum) as well as to facilitate measurements across aggregated set of search results, rather than separately for each result list. Finally, we showcase the metric through a simulated audit of a hiring service, an online dating service, and a search engine. We show that knowing the usage patterns of the particular service is crucial in determining the fairness of its results---depending on the attention distribution function, the same list of results can appear biased both in favor and against a protected group.
READ FULL TEXT