Accuracy, Interpretability, and Differential Privacy via Explainable Boosting

We show that adding differential privacy to Explainable Boosting Machines\n(EBMs), a recent method for training interpretable ML models, yields\nstate-of-the-art accuracy while protecting privacy. Our experiments on multiple\nclassification and regression datasets show that DP-EBM models suffer\nsurprisingly little accuracy loss even with strong differential privacy\nguarantees. In addition to high accuracy, two other benefits of applying DP to\nEBMs are: a) trained models provide exact global and local interpretability,\nwhich is often important in settings where differential privacy is needed; and\nb) the models can be edited after training without loss of privacy to correct\nerrors which DP noise may have introduced.\n