FACET: Counterfactual Explanation Analytics

VanNostrand, Peter

Etd

FACET: Counterfactual Explanation Analytics

Public Deposited

In recent years, machine learning systems have become ubiquitous in domains such as finance, hiring, and healthcare where an undesired outcome generated by a model can have serious ramifications for the user. This has created a rising demand for AI-powered systems capable of explaining what led to the given decision and identifying changes which when applied to a user's instance alter the model's decision to their desired outcome. To meet this need, we propose FACET, the first explanation analytics system which supports a user in interactively refining counterfactual explanations for decisions made by tree ensembles. As FACET's foundation, we design a novel type of counterfactual explanation called the counterfactual region. These regions concisely describe portions of the feature space over which the user is guaranteed to achieve their desired outcome, regardless of variation in the exact feature values of their instance. We coin this property as explanation robustness and identify it as critical for the practical application of counterfactuals. By efficiently processing a rich set of explanation analytic queries, FACET empowers the user to identify personalized counterfactual regions which account for their real-world circumstances and best meet their personal preferences. We develop a novel bit-vector based counterfactual region explanation index, called COREX, which accelerates FACET's query execution to allow for near real-time exploration of explanations regardless of model complexity. We evaluate FACET against state-of-the-art explanation techniques on five publicly available benchmark datasets and demonstrate that FACET generates explanations of similar quality in an order of magnitude less time -- enabling critical human-in-the-loop interactions.

Creator