ExplainabilityTransparencyTrustworthinesscs.LG
Generating robust counterfactual explanations
Victor Guyomard, Françoise Fessant, Thomas Guyet, Tassadit Bouadi, Alexandre Termier
April 24, 2023
ECML/PKDD
22 citations
Key Takeaways
-
Counterfactual explanations need to be robust: The paper identifies that current counterfactual explanations—which tell users what changes would alter a model's decision—can be fragile. Small variations in the input data can cause these explanations to become invalid, undermining their usefulness for understanding AI decisions.
-
There's a fundamental trade-off to manage: The research reveals tension between two desirable qualities: keeping counterfactuals close to the original situation (so they're realistic and actionable) while making them robust to input changes. The CROCO framework addresses this by explicitly managing this trade-off rather than ignoring it.
-
Robustness guarantees matter for trust: By providing users with "minimal robustness" guarantees, the framework helps build confidence that explanations won't suddenly become misleading if real-world data shifts slightly—an important consideration for deploying AI systems responsibly in practice.
-
The problem is particularly acute for tabular data: The empirical evaluation focused on tabular datasets (spreadsheet-like data), suggesting this robustness challenge is especially relevant for common business applications like credit decisions, hiring, and loan approvals where counterfactual explanations are frequently used.
Abstract
Counterfactual explanations have become a mainstay of the XAI field. This particularly intuitive statement allows the user to understand what small but necessary changes would have to be made to a given situation in order to change a model prediction. The quality of a counterfactual depends on several criteria: realism, actionability, validity, robustness, etc. In this paper, we are interested in the notion of robustness of a counterfactual. More precisely, we focus on robustness to counterfactual input changes. This form of robustness is particularly challenging as it involves a trade-off between the robustness of the counterfactual and the proximity with the example to explain. We propose a new framework, CROCO, that generates robust counterfactuals while managing effectively this trade-off, and guarantees the user a minimal robustness. An empirical evaluation on tabular datasets confirms the relevance and effectiveness of our approach.