2021 · 87 citations · 44 references
ComputingComputer ScienceInformation SystemsArtificial IntelligenceData AnalyticsMachine LearningData InterpretationExplainable Machine LearningExplanationFraud DetectionFraud InvestigationExplainabilityExplainable Artificial IntelligenceExplanation Quality EvaluationHuman-centered Explainable Artificial IntelligenceLime ExplanationsModel InterpretabilityPost Hoc ExplanationsPractical Effects
Explainable AI methods are often designed for properties like fidelity or interpretability, yet they are rarely evaluated for their real impact on decision‑making, risking degraded overall system performance. This study proposes XAI Test, an application‑grounded evaluation method to isolate the effect of different explanation levels on end‑user decisions. Using XAI Test, we evaluated LIME, SHAP, and TreeInterpreter on a fraud‑detection task with real data and analysts, progressively adding data, model scores, and explanations in three stages. Statistical analysis shows that explanations generally worsen performance: Data‑Only yields highest accuracy and slowest decisions, while all explainers improve over Data + Score but still underperform Data‑Only; LIME is least preferred due to low explanation variability.
There have been several research works proposing new Explainable AI (XAI) methods designed to generate model explanations having specific properties, or desiderata, such as fidelity, robustness, or human-interpretability. However, explanations are seldom evaluated based on their true practical impact on decision-making tasks. Without that assessment, explanations might be chosen that, in fact, hurt the overall performance of the combined system of ML model + end-users. This study aims to bridge this gap by proposing XAI Test, an application-grounded evaluation methodology tailored to isolate the impact of providing the end-user with different levels of information. We conducted an experiment following XAI Test to evaluate three popular post-hoc explanation methods -- LIME, SHAP, and TreeInterpreter -- on a real-world fraud detection task, with real data, a deployed ML model, and fraud analysts. During the experiment, we gradually increased the information provided to the fraud analysts in three stages: Data Only, i.e., just transaction data without access to model score nor explanations, Data + ML Model Score, and Data + ML Model Score + Explanations. Using strong statistical analysis, we show that, in general, these popular explainers have a worse impact than desired. Some of the conclusion highlights include: i) showing Data Only results in the highest decision accuracy and the slowest decision time among all variants tested, ii) all the explainers improve accuracy over the Data + ML Model Score variant but still result in lower accuracy when compared with Data Only; iii) LIME was the least preferred by users, probably due to its substantially lower variability of explanations from case to case.
44
Leo Breiman · Machine Learning · 2001 · 126.2K citations · Full text
A Simple Sequentially Rejective Multiple Test Procedure
Sture Holm · Scandinavian Journal of Statistics · 1979 · 21.8K citations
Marco Túlio Ribeiro, Sameer Singh, Carlos Guestrin · 2016 · 15.2K citations
Advances in Neural Information Processing Systems 29
Onur Teymur, Kostas Zygalakis, Ben Calderhead · 2016 · 13.5K citations