Reviews: Fairness Through Computationally-Bounded Awareness

The paper builds on the influential "Fairness Through Awareness" paper of Dwork et al., which provided a framework for learning non-discriminatory classifiers by requiring the the classifier to treat similar individuals similarly. This similarity was defined by a (hypothetical) task-specific metric for determining the degree to which individuals are similar with respect to the classification task at hand. This paper extends the results to the more realistic scenario when the entire metric is not known to the learner, while relaxing the requirement that all pairs of similar individuals be treated similarly. The paper provides both learning algorithms and hardness results for this model. Fairness in machine learning is still a very young field, with many competing models and definitions, and more questions than answers.