Measuring Discrimination in Decision-Making Algorithms: An Approach Based on Causal Inference

Machine learning algorithms are routinely used for business decisions that may directly affect individuals in various contexts, such as credit scoring, employment, and criminal justice. When such algorithms are used in the decision process, their behavior concerning discrimination depends on the information it is given, and discrimination may occur unconsciously or explicitly based on sensitive attributes. Statistical tools and methods are then required to handle such potential biases. We propose to exploit the Coarsened Exact Matching (CEM) algorithm to measure discrimination against a protected group to be used in data pre-processing for discrimination detection and removal. Experiments are conducted to test the proposed methodology on real data and a comparison with related work is also discussed.