New fairness measure applied to a learning to rank method

Abstract

Nowadays, the need to organize data and enhance its accessibility in a systematic way is ubiquitous. However, data is not free from bias and preconceptions, and the inherent data bias may negatively influence data-driven algorithms. Consequently, the information represented may tend to favour certain groups over others, thereby perpetuating discrimination against so-called protected groups. This is also particularly evident in the field of learning to rank (LtR), in which LtR algorithms are trained to rank a set of items represented as feature vectors. Numerous studies have been conducted in recent years on fairness management for machine learning algorithms, with the objective of reducing the effects of data biases on the trained models. In this thesis, we focus on the relationship between LtR and fairness. We modify an LtR framework, LambdaMART, whose original goal is to optimize NDCG@k, by adding a group-based fairness measure to optimize, called Normalised Discounted Difference (rND). Following an initial study focusing on LtR and fairness, various methodologies for combining the two metrics and their application on two real datasets will be proposed and evaluated.