Unfairness and Discrimination in Machine Learning: how adding the NPI regularizer could affect classification outputs

In the last decades, machine learning has played an ever-growing role in supporting decision-making processes, enhancing their speed and efficiency. However, recent researches highlighted the presence of some risks regarding the safety of these processes in terms of possible violations of fundamental human rights. For example, some papers showed that discrimination and unfairness may be embedded in the predictions coming from the use of algorithms. At the moment, no unifying solution has been found to solve this problem but many different approaches have been proposed. Given the high contextuality of the matter this thesis does not try to offer one global answer to the issue, but it rather focuses on a specific method already proposed in the literature, discovering its advantages and disadvantages, extending its applicability, while giving first a brief overview of the most accepted high-level definitions and methods. The analyzed approach is the one developed in the article called Fairness-Aware Classifier with Prejudice Remover Regularizer written by Kamashima, Akaho, Asoh and Sakuma. Shortly, it consists in adding a regularizer to the objective function of a classification model to decrease the predictions unfairness that arises from a sensitive binary variable. However, the presence of just one sensitive binary variable is rather unrealistic considering the complexity of the world, this is why the extension of this method to a categorical variable with more then two categories, to a continuous variable and to more than one binary variable will be presented. The optimization of both the original method and its extensions has been implemented in R using a code created only for the purpose of this thesis. Defining what is ethical and what is not has always been a difficult task in many fields and machine learning is no exclusion. This is why raising awareness on the potential harm that the application of algorithms could do to people is an important first step that this paper hopes to contribute to, while also continuing the discussion on such an important problem that due to its circumstantiality will most likely always need a key human oversight role in its development and understanding.