FAIR: Fair adversarial instance re-weighting

•We merge reweighing and adversarial approaches for mitigating bias in machine learning models, while keeping the best from both.•The proposed method can provide interpretable information about fairness of individual instances.•We provide theoretical analysis of properties of adversarial re-weighting.•We explore several variants of instance weight estimation including probabilistic ones.•We evaluate the method on four different real-world datasets, compared to the state-of-the-art techniques, and provide qualitative analysis.