ScaDS.AI Dresden/Leipzig announces and welcomes you to join its public colloquium session on Tuesday, May 27, 2025 at 1:15 pm CEST. The colloquium takes place at seminar room “Zwenkauer See” at ScaDS.AI Dresden/Leipzig (details below) and parallel online (link to Zoom session).

Conditional Feature Importance with Generative Modeling

A core question in explainable machine learning is: which features are truly important for a model’s predictions? Importantly, feature importance is often conditional—the effect of one feature may depend on the values of others. Estimating this conditional importance involves evaluating model performance under feature perturbations that preserve the joint structure of the data. This requires generating synthetic but realistic samples from conditional feature distributions, which is technically nontrivial, especially for mixed-type tabular data.
In this talk, I’ll present cARFi, a model-agnostic method that leverages adversarial random forests (ARF) to approximate these conditional distributions. ARF provides a lightweight, distribution-aware generative model that avoids the complexity of deep generative networks. By permuting features using ARF-sampled values, cARFi offers a practical and flexible framework for measuring conditional (and marginal) feature importance. It also enables statistical testing to assess the significance of observed importance scores.

About Prof. Dr. Markus Loecher

Prof. Dr. Markus Loecher is a Professor of Mathematics and Statistics at the Berlin School of Economics and Law (HWR Berlin), where he has held a faculty position since 2011. His research spans machine learning, explainable artificial intelligence (XAI), data visualization, and sequential decision-making.
Before entering academia, he worked as a principal and lead data scientist at several analytics firms in the United States. He holds a Ph.D. in Physics from Ohio University and an M.S. in Statistics from Rutgers University.
Prof. Loecher is the inventor or co-inventor of eight patents in the areas of machine learning and sensor analytics. He is also the author of several widely used R packages, including RgoogleMaps (integrating web map tiles into R), bandit (tools for multi-armed bandits and Thompson sampling), and rfVarImpOOB (unbiased variable importance for random forests). He has published more than 35 peer-reviewed articles in international journals and is the author of a book on nonlinear stochastic phenomena.

Location

ScaDS.AI Dresden/Leipzig
Löhrs Carré, Humboldtstrasse 25, 04105 Leipzig
3rd floor, large seminar room (A 03.07 “Zwenkauer See”)