Explanations that precisely address the question

Congratulations 🎉 to BIFOLD researchers Simon Letzgus, Klaus-Robert Müller and Grégoire Montavon, whose publication: XpertAI: uncovering regression model strategies for sub-manifolds won the BestPaperAward at the 3rd World Conference on eXplainable Artificial Intelligence.The Conference took place July 09-11, 2025 in Istanbul, Turkey.
This paper explores Explainable AI for machine learning models that solve regression tasks. Regression tasks involve predicting real-valued quantities such as price, age, quality, efficiency, etc. In this paper, the researchers emphasize the importance of carefully specifying "what" to explain in regression tasks. For example, when predicting perceived wine quality based on its chemical properties and seeking an explanation of how the model makes predictions, the answers to "What distinguishes an excellent wine from a good wine?" and "What sets it apart from a bad wine?" can differ significantly. Technically, the paper introduces a novel, flexible methodology in which the ML model is deconstructed into a set of "range experts", each focusing on a different narrow output range. Users can then query this pool of range experts by combining them linearly to obtain explanations that precisely address their questions.
BIFOLD researchers contributed on several levels to this leading conference on XAI. Below is an overview of BIFOLD’s contributions:
Debugging and improving models with XAI. Christopher Anders (Riken / BIFOLD Alumni), Sebastian Lapuschkin (Fraunhofer HHI), Frederik Pahde (Fraunhofer HHI), Maximilian Dreyer (Fraunhofer HHI), Lorenz Linhardt (BIFOLD).
[Track]
Actionable Explainable AI. Grégoire Montavon (Charité - Universitätsmedizin Berlin / BIFOLD), Lorenz Linhardt (BIFOLD), Caroline Petitjean ( Université de Rouen), Gian Antonio Susto (University of Padua).
[Track]
Steering committee: Grégoire Montavon (Charité - Universitätsmedizin Berlin / BIFOLD)
Late-breaking work & demos committee: Lorenz Linhardt (BIFOLD), Oliver Eberle (BIFOLD)