Counterfactual Explanations for Data-Driven Decisions.

Users' lack of understanding of systems that use predictive models to make automated decisions is one of the main barriers for their adoption. We adopt the increasingly accepted view of a counterfactual explanation for a system decision: a set of the system inputs that is causal (meaning that removing them changes the decision) and irreducible (meaning that removing any subset of the inputs in the explanation does not change the decision). We generalize previous work on counterfactual explanations in three ways: we explain system decisions rather than model predictions; we do not enforce any specific method for removing inputs, and our explanations can incorporate inputs with arbitrary data structures. We also show how model-agnostic algorithms can be tweaked to find the most useful explanations depending on the context. Finally, we showcase our approach using a real data set to illustrate its advantages over other explanation methods when the goal is to understand system decisions better. [ABSTRACT FROM AUTHOR]