Reducing Disparate Exposure in Ranking: A Learning To Rank Approach
In this paper we consider a ranking problem in which we would like to order a set of items by utility or relevance, while also considering the visibility of different groups of items. To solve this problem, we adopt a supervised learning to rank approach that learns a ranking function from a set of training examples, which are queries and ranked lists of documents for each query. We consider that the elements to be ranked are divided into two groups: protected and non-protected. Following long-standing empirical observations showing that users of information retrieval systems rarely look past the first few results, we consider that some items receive more exposure than others. Our objective is to produce a ranker that is able to reproduce the ordering of the training set, which is the standard objective in learning to rank, but that additionally gives protected elements sufficient exposure, compared to non-protected elements. We demonstrate how to describe this objective formally, how to achieve it effectively and implement it, and present an experimental study describing how large differences in exposure can be reduced without having to introduce large distortions in the ranking utility.
READ FULL TEXT