Combining white box models, black box machines and human interventions for interpretable decision strategies

Abstract Granting a short-term loan is a critical decision. A great deal of research has concerned the prediction of credit default, notably through Machine Learning (ML) algorithms. However, given that their black-box nature has sometimes led to unwanted outcomes, comprehensibility in ML guided decision-making strategies has become more important. In many domains, transparency and accountability are no longer optional. In this article, instead of opposing white-box against black-box models, we use a multi-step procedure that combines the Fast and Frugal Tree (FFT) methodology of Martignon et al. (2005) and Phillips et al. (2017) with the extraction of post-hoc explainable information from ensemble ML models. New interpretable models are then built thanks to the inclusion of explainable ML outputs chosen by human intervention. Our methodology improves significantly the accuracy of the FFT predictions while preserving their explainable nature. We apply our approach to a dataset of short-term loans granted to borrowers in the UK, and show how complex machine learning can challenge simpler machines and help decision makers.