Group Fairness

Assessing whether a model is discriminatory, or not, is a complex problem. As in Chap. 3 , where we discussed global and local interpretability of predictive models, we start with some global approaches (the local ones will be discussed in Chap. 9 ), also called “group fairness,” comparing quantities between groups, usually identified by sensitive attributes (e.g., gender, ethnicity, age, etc.). Using the formalism introduced in the previous chapters, y denotes the variable of interest, \(\widehat {y}\) or \(m(\boldsymbol {x})\) denotes the prediction given by the model, and s the sensitive attribute. Most concepts are derived from three main principles: independence ( ), separation ( conditional on y), and sufficiency (( ) conditional on \(\widehat {y}\) ). We review these approaches here, linking them while opposing them, and we implement metrics related to those notions on various datasets.