FAIR: Fair adversarial instance re-weighting

With growing awareness of societal impact artificial intelligence, fairness has become an important aspect machine learning algorithms. The issue is that human biases towards certain groups population, defined by sensitive features like race and gender, are introduced to the training data through collection labeling. Two directions ensuring research have focused on (i) instance weighting in order decrease more biased instances (ii) adversarial construct representations informative target variable, but uninformative attributes. In this paper we propose a Fair Adversarial Instance Re-weighting (FAIR) method, which uses learn function ensures fair predictions. Merging two paradigms, it inherits desirable properties from both interpretability reweighting end-to-end trainability training. We four different variants method and, among other things, demonstrate how can be cast fully probabilistic framework. Additionally, theoretical analysis FAIR models’ provided. compare models ten related state-of-the-art able achieve better trade-off between accuracy unfairness. To best our knowledge, first model merges approaches means provide interpretable information about individual instances.