Prospective Explanations: An Interactive Mechanism for Model Understanding
Proceedings of the NeurIPS 2021 Competitions and Demonstrations Track, PMLR 176:273-277, 2022.
We demonstrate a system for prospective explanations of black box models for regression and classification tasks with structured data. Prospective explanations are aimed at showing how models function by highlighting likely changes in model outcomes under changes in input. This in contrast to most post-hoc explanability methods, that aim to provide a justification for a decision retrospectively. To do so, we employ a surrogate Bayesian network model and learn dependencies through a structure learning task. Our system is designed to provide fast estimates of changes in outcomes for any arbitrary exploratory query from users. Such queries are typical partial, i.e. involve only a selected number of features, the outcomes labels are shown therefore as likelihoods. Repeated queries can indicate which aspects of the feature space are more likely to influence the target variable. We demonstrate the system from a real-world application from the humanitarian sector and show the value of bayesian network surrogates.