Achieving non-discrimination in prediction

Discrimination-aware classification is receiving an increasing attention in the data mining and machine learning fields. The data preprocessing methods for constructing a discrimination-free classifier remove discrimination from the training data, and learn the classifier from the cleaned data. However, there lacks of a theoretical guarantee for the performance of these methods. In this paper, we fill this theoretical gap by mathematically bounding the probability that the discrimination in predictions is within a given interval in terms of the given training data and classifier. In our analysis, we adopt the causal model for modeling the mechanisms in data generation, and formally defining discrimination in the population, in a dataset, and in the prediction. The theoretical results show that the fundamental assumption made by the data preprocessing methods is not correct. Finally, we develop a framework for constructing a discrimination-free classifier with a theoretical guarantee.