Trustworthy Feature Importance Measures : New Findings

Place : Room 3.216 - Grand Paris Campus & online 
Time : 12.00-01.30 PM

Speaker : Emanuele Borgonovo, Department of Decision Sciences and Bocconi Institute for Data Science and Analytics, Bocconi University, Milan, Italy.

Abstract : The determination of feature importance is essential for explainability. However, methods based on unrestricted permutations deliver incorrect insights due to extrapolation errors. This is a problem that affects essentially all non-trivial variable importance metrics to date, leading them to make untrustworthy estimates of the importance of variables. We introduce an effective solution to address the problem -- a quantile transformation. By permuting \textit{quantiles} rather than \textit{values}, we ensure the transformed values are never impossible. We apply the transformation to the calculation of conditional model reliance with a Gaussian transformation of the features, to obtain an index that includes only the unique information from the feature of interest. Second, we propose a strategy that combines Knockoffs for the generation of the new data and a Gaussian transformation. We establish the theoretical connections of the new variable importance measures with total indices under a quadratic loss. We examine the performance of the new indices through extensive numerical experiments, using analytical test cases with increasing dimensionality, as well as four well-known datasets with alternative feature dimensionality and sample size: Boston Housing, Bike Sharing, Spambase and the recently introduced Voter-Name-Ethnicity Prediction dataset. Results reveal that the proposed permutation strategies successfully decrease extrapolation and are much more trustworthy for estimating the importance of variables than their unrestricted counterparts.

Biography : Emanuele Borgonovo is Full Professor and Director at the Department of Decision Sciences of Bocconi Universit and Research Affiliate at the Massachusetts Institute of Technology, Department of Nuclear Science and Engineering. He holds a Ph.D. in Nuclear Science and Engineering from MIT. He is co-editor-in-chief of the European Journal of Operational Research, advisor to the Springer International Series in Operations Research and Management Science, Past-President of the INFORMS Decision Analysis Society, member of the Scientific Committee of the Silvio Tronchetti Provera Foundation and Co-Chair of the Uncertainty Analysis Committee of the Safety and Reliability Association.  He is the recipient of several national and international awards, among which the 2017 Chinese Academy of Sciences President’s Visiting Professor Fellowship, the 2015 IBM Faculty Award, the honorary memberships of “, The Scientific Research Society of North America”, “The Honorary Society of the American Nuclear Society,” the “McCormack fellowship of the Westinghouse Corporation” as well as several best paper and excellence in refereeing awards. He is also a member of the editorial boards of numerous journals among which Risk Analysis, Decision Analysis, Reliability Engineering and System Safety, the Journal of Financial Data Science, and he is associate editor of the Journal of Risk and Reliability. He has published over 100 scientific papers and worked in international research projects with institutions such as DARPA, the Idaho National Laboratory, the DOE, the US Nuclear Regulatory Commission, Electricité de France, Charles Rivers Analytics, The Marie Technimont Group, the Campari Group, etc.. The Differential Importance Measure, a sensitivity technique he introduced in 2000, has been inserted in NASA’s Guide for Risk Managers and Practitioners. He is author of the book: “Sensitivity Analysis: An Introduction for the Management Scientist”, http://www.springer.com/gp/book/9783319522579.

Pour plus d'informations, merci de contacter Margherita Pagani : margherita.pagani@skema.edu