All of the Fairness for Edge Prediction with Optimal Transport

We live in a world where an increasing number of decisions, with major societal consequences, are made or at least supported by algorithms that diligently learn the patterns from a training sample and gain their discriminating ability by identifying the key attributes correlated with the desired output. These attributes, however, can represent sensitive information that, in its turn, can lead to a significant bias in model's predictions when deployed on a previously unseen sample. For instance, when building a recommendation system supporting a recruitment company in finding a potential candidate suitable for their clients' needs, one would expect its recommendations to be independent from the gender or the ethnicity of the considered individuals. In practice, however, the training sample used to learn the model may have been collected in a biased manner with an unequal number of successive outcomes between the genders and/or ethnic groups. The recommendations of the learned model in this case will tend to follow the learned pattern thus reinforcing the already existent bias. Research works aiming at identifying and correcting such inductive bias form the core of the algorithmic fairness field, a scientific area that is constantly gaining more and more attention from the machine learning and data mining communities nowadays. Algorithmic fairness methods are traditionally divided into one of the three following categories: (i) pre-processing methods that repair the original data to remove the bias, ii) methods that integrate fairness constraints or penalties in a given learning algorithm and iii) post-processing methods that debias directly the model's output. First family of methods can be further divided into two subfamilies where the first one corrects the input raw data to ensure that the inference of the sensitive attribute is impossible, regardless of the learning algorithm (e.g.