The benefits of machine-learned models only apply if the training and the test data come from the same distributions. In many application scenarios, distribution shifts outdate the learned models - often surprisingly quickly. Model monitoring is required not to build your application on wrong assumptions. In the best world, up-to-date validation data would allow for fully informed model monitoring. In realistic scenarios, such up-to-date validation data is frequently missing, requiring methods for unsupervised model monitoring.
Established unsupervised model monitoring approaches check for shifts of input data or predicted outcomes, but both lead to too many false positives and negatives.
Attribution approaches (e.g. Shapley Values) evaluate how a learned model picks up or ignores features of an input data record. An explanation distribution, thus, represents how a learned model handles data. The core novelty of our approach is to introduce this notion of explanation distribution and check for shifts of explanation distributions from training to test data, i.e., monitor a potential explanation shift.
Check out the paper
Carlos Mougan, Klaus Broelemann, Gjergji Kasneci, Thanassis Tiropanis, Steffen Staab. Explanation Shift: How Did the Distribution Shift Impact the Model? In Transactions on Machine Learning Research, 2025. https://openreview.net/forum?id=MO1slfU9xy