Machine learning models tested under shifts in input data source
Counterfactual Model Testing Under Distribution Shift
Artificial IntelligenceMachine Learning
Summary
Machine learning models often perform worse when the data they see changes over time. These changes can come from either actual shifts in the population or problems with how data is collected. The authors propose a way to test models by recreating these changes separately to understand how each affects model performance. This helps estimate current accuracy and decide if fixing data problems can improve results. They show with experiments that treating these sources differently leads to better testing and repair decisions.
What this means in practice
- •For data engineering teams: Support monitoring and prioritizing fixes for data issues that affect deployed model performance.
- •For business analytics teams: Estimate how population changes impact predictive models to better understand model reliability over time.
Authors
Joran Leest, Ilias Gerostathopoulos, Claudia Raibulet, Patricia Lago
Abstract
Deployed machine learning (ML) models can degrade as the distribution of their input data shifts. Such shifts can result from changes in the population or from data integrity issues in the pipeline that materializes model features, two sources that operate through different mechanisms. Because labels are often absent or delayed, it is common practice to track shifts in the data instead, but a detected shift does not indicate how it affects the model. We formulate this problem as online model testing under distribution shift. Testing requires reproducing the shift on a controlled baseline with known labels. Since the same shift can come from either source, doing so requires knowing which mechanism changed. We propose a methodology that reproduces observed changes in both mechanisms to estimate current performance and the potential gain from repairing observed data integrity issues. Experiments on real-world datasets with simulated faults show that separating the two sources generally improves both estimates, supporting this formulation. A churn prediction case illustrates how these tests can support monitoring and repair prioritization