End-to-end Learning for Fair Ranking Systems

Current approaches to fairness in learning-to-rank systems rely on using a loss function representing a weighted combination of The learning-to-rank problem aims at ranking items to maximize expected task performance and fairness. This strategy is effective exposure of those most relevant to a user query. A desirable property in improving the fairness of predicted rankings on average, but has of such ranking systems is to guarantee some notion of fairness three key shortcomings: (1) The resulting rankings, even when fair among specified item groups. While fairness has recently been considered in expectation across all queries, can admit large fairness disparities in the context of learning-to-rank systems, current methods for some queries. This aspect may contribute to exacerbate the richget-richer cannot provide guarantees on the fairness of the proposed ranking dynamics, while giving a false sense of controlling the policies. This paper addresses this gap and introduces Smart Predict system's disparate impacts.