Fairness-Guided Pruning of Decision Trees

Decision tree learning is a popular machine learning technique, in particular for applications where the interpretability of the predictions is crucial—such as applications in health and financial domains. When dealing with a large dataset, decision tree algorithms poten- tially generate a complex model that overfits the training data. In such cases, it becomes challenging for decision trees to maintain an interpretable structure while also identifying potential biases in predictions, raising concerns about poor performance on unseen data. To address this issue, decision trees can be pruned to reduce their size, as a result enhancing interpretability and improving predictive accuracy on new instances through the use of simplified models. However, traditional pruning methods typically focus solely on predictive accuracy, which may inadvertently increase model bias and negatively affect fairness. Moreover, current post-processing fairness techniques often aim to reduce discrimination by modifying the tree's labels without considering the complexity inherent in large trees. To address these challenges, we propose a novel fairness-guided pruning strategy for decision trees that improves both fairness and interpretability. Computational experiments com- paring our proposed strategy with existing methods demonstrate that our fairness-guided pruning achieves a good accuracy-fairness trade-off overall: small reductions in predictive accuracy are associated with improvements in fairness while simplifying the decision tree structure at the same time.