Causal Discovery and Causal Inference Based Counterfactual Fairness in Machine Learning

Abstract

The fairness problem arouses attention in machine learning. One problem with traditional counterfactual fairness is the assumed causal models are constrained by prior knowledge. We propose a framework named Structural Causal Fairness Framework (SCFF) to achieve counterfactual fairness without assumptions like previous works. To correct observations adversely affected by the sensitive attributes, we follow the objectives of fair sampling and construct structural causal models based on causal discovery and causal inference. Experiments show our framework generates competitive results on both counterfactual fairness level and prediction accuracy compared with the other three baselines. More importantly, our framework is all based on data and has good generalization on machine learning problems.