Fairness improvement for black-box classifiers with Gaussian process

Abstract

In many real-world applications, it is the fairness, not the accuracy, of a machine learning (ML) classifier that is the crucial factor. Post-processing approaches are widely considered as successful tools to improve the fairness of black-box ML classifiers. These aim to learn a relabeling function to modify initial predicted labels provided by a pre-trained “unfair” classifier, resulting in fair classification on a given test set. However, many post-processing methods require a training set with true labels to learn the relabeling function. To the best of our knowledge, there have been only two methods that learn the relabeling function without requiring the true labels of training samples. However, both of these methods require access to the predictions of the pre-trained classifier when performing on the test set, even after they learned the optimal relabeling function, and neither offers theoretical guarantees on the trade-off between accuracy loss and fairness improvement. In this paper, we propose a novel post-processing method based on Gaussian process (GP). We first train a GP with unlabeled samples, and use its posterior mean function to approximate the predictions of the pre-trained classifier. We then adjust the mean function (i.e. the relabeling function) to achieve two goals: (1) maximize the fairness and (2) minimize the difference between the relabeling function and the pre-trained classifier. By doing this, our method can improve fairness while maintaining high accuracy. We provide a theoretical analysis to derive an upper bound on accuracy loss for our method. We demonstrate our method on four real-world datasets, comparing with state-of-the-art baselines, to demonstrate its ability to achieve both fairness and accuracy.