ANR Project: GATSBII (2025-2029)

GAme Theory and Statistical estimation Bring Importance Measures and Interpretability



Project summary

The extensive use of machine learning (ML) models in data-based artificial intelligence (AI) systems --especially those submitted to new European regulations-- require explanations, to serve the intelligibility and auditability of these AI systems. Explainability in AI (XAI) is linked to the ability, for a human mind, to understand and communicate about decisions proposed by ML models. One of the primary challenges in XAI is to understand the influence of the features ("inputs") on the predicted variables (“outputs”) and to provide global interpretability diagnostics. Strong connections between XAI and Global Sensitivity Analysis (GSA) of model outputs have been recently highlighted . Indeed, GSA aims at studying how the uncertainty in a model output can be apportioned to different sources of uncertainty in its inputs. In the last decades, applied mathematicians have developed several GSA indices (also called importance measures (IM)) to quantify the relative importance of the inputs on the output while heuristic methods are mainly used in XAI.

In GSA and XAI, several fundamental open issues remain. How to consider the statistical dependency between inputs, and in particular, how to distinguish its effects from the inputs' interaction effects? How to estimate indices when one cannot choose a priori the input design of experiments? How can we estimate online the indices? How can we deal with high dimensional models? These open issues are even more relevant in a industrial where the input variables are rarely independent and the design of experiment can not be chosen. The project aims to face these challenging questions by gathering together statisticians specialists/practitioners of GSA/XAI and researchers in Cooperative Game Theory by exploring several tracks (hierarchical models and tools from game theory to deal with dependency, stochastic algorithm to built on line estimators, ... )

People

Internal meetings

2025, June, 16, Paris

2025, January, 30-31, Toulouse

Master internship, PhD thesis and Post-doctoral offers

Publications

  • Thierry Klein, Agnès Lagnoux, Thi Mong Ngoc Nguyen, Paul Rochet, Asymptotic efficiency for Sobol' and Cramér-von Mises indices under two designs of experiments: Submitted Juillet 2025.https://hal.science/hal-04655042

    Communications

      • L. Clouvel, B. Iooss and V. Chabridon, Variance-based importance measures for high-dimensional linear model via Johnson indices: Insights and comparisons, 11th International Conference on Sensitivity Analysis of Model Output, April 23–25 2025, Grenoble, France - SAMO book of abstracts
      • A. Marrel, G. Sarazin, B. Iooss and A. Spagnol, Overcoming challenges in sensitivity analysis for complex models in nuclear and renewable energy applications, 11th International Conference on Sensitivity Analysis of Model Output, April 23–25 2025, Grenoble, Francerenewable energy applications - SAMO book of abstracts

      Software