MMDFast: Efficient Understanding of Machine Learning Model Mispredictions

Mispredictions in machine learning models pose significant risks, particularly in safety-critical systems where reliability is paramount. Consequently, debugging these errors is vital. While existing methods attempt to generate rules to explain mispredictions, they suffer from prohibitive computational costs, often rendering them unusable on large datasets. We propose MMDFast, a novel approach that overcomes these limitations by (i) utilizing a lightweight auxiliary model to identify and focus only on the most influential features correlated with mispredictions, and (ii) replacing exhaustive rule search with optimized decision tree learning. Our evaluation across 11 diverse, real-world datasets shows that MMDFast not only significantly improves the accuracy of misprediction explanations but also drastically reduces computational time. MMDFast [ECG25] was accepted for publication at the International Conference on Automated Software Engineering in the research track (ASE 2025).