Generic Framework for Explainable Fairness in Machine Learning

Given that machine learning has been used in a wide range of applications, it is crucial to mitigate unfairness during training process. This work concentrates on the well-known idea of individual fairness, which holds that similar individuals should be treated similarly in similar tasks assigned to them. Although individual fairness can be enhanced during the model training process (in-processing), pre-processing the data is still a important operation before model training. Particularly, we demonstrate that individual records reassigned within the k-nearest-neighbours are a powerful pre-processing method for enhancing individual fairness. Our proposed framework, called iShare, addresses the problem of optimising labels given a limit to the number of unfairness violations. We first prove that the problem is NP-hard, and then propose an approximate linear programming algorithm and provide theoretical guarantees on how close its result is to the optimal solution. Moreover, we develop a new method to enable the linear programming solution close to optimal within the constraints. Experimental results show that our iShare framework significantly outperforms other pre-processing benchmarks with regards to individual fairness score and accuracy on unseen data on real-world datasets