|
|
| |
INVITED SEMINAR SERIES
Chair Risk and Resilience of Complex Systems | LGI | CentraleSupélec |
| |
|
|
|
|
| |
|
|
| |
Reliable Explainable AI: From Trustworthy Explanations to Variable Importance for Unsupervised Learning
Dr. Xuefei LU
Associated Professor, SKEMA Centre for Analytics and Management Science
SKEMA Business School, France
|
|
|
| |
|
|
| |
|
Time: Friday, July 03, 2026, 10:00 – 12:00 am CET
|
|
|
| |
|
|
| |
Abstract Explainable artificial intelligence (XAI) has become increasingly important for ensuring transparency, regulatory compliance, and informative decision making. However, the robust and trustworthiness of existing explanation methods remains an open challenge. Popular approaches such as SHAP, LIME, and permutation-based techniques often rely on synthetic data points that may lie outside the data manifold, producing explanations based on unrealistic or infeasible inputs. Such extrapolation can undermine the trustworthiness of model explanations, particularly in high-stakes applications.
This talk presents recent advances toward more reliable explainable AI from two complementary perspectives. First, a systematic framework is proposed to perform sanity checks on machine learning explanations. By combining statistical tests with visual diagnostics, the framework provides practical tools for assessing explanation reliability and enhancing the faithfulness of a wide range of XAI methods.
The talk then considers the problem of variable importance in clustering, where the absence of ground-truth labels makes interpretation particularly challenging. A novel global sensitivity analysis framework based on optimal transport is presented, introducing a Wasserstein-based variable importance measure that quantifies how each variable influences the global clustering structure. The methodology is applicable to both hard and soft clustering algorithms. |
|
|
| |
|
|
|
|
|
Xuefei LU is an Associate Professor at SKEMA Business School, France. She holds a PhD in Statistics from Bocconi University (Italy) and has previously worked as a post-doctoral researcher at Politecnico di Milano (Italy) and as a lecturer at the University of Edinburgh (UK).
Her research focuses on explainable artificial intelligence, statistical machine learning, and uncertainty quantification, with publications appearing in journals such as Operations Research, European Journal of Operational Research, Decision Analysis, Reliability Engineering & System Safety, etc.
|
|
|
|
|
| |
|
Her works have received international awards, among which the 2023 INFORMS Clemen-Kleinmuntz award and Runner-Up of 2024 SAS Data Mining Best Paper Award.
She is actively involved in academic service, serving on the editorial board of the European Journal of Operational Research and on the scientific committee of the Sensitivity Analysis of Model Output (SAMO) community. She is also a council member of the INFORMS Decision Analysis Society and a frequent reviewer for top-tier journals. .
|
|
|
|
|
|
|
|
Chaire on risk and resilience of complex systems
Laboratoire Génie Industriel (LGI)
CentraleSupélec
3 rue Joliot-Curie F-91192 Gif-sur-Yvette France
|
|
|
|
|
|
