Settles the long-standing practitioner debate over whether to use training or holdout data for interpreting black-box models with PD/ALE plots.
arXiv · March 17, 2026 · 2603.15057
The Takeaway
It shows that the bias introduced by using training data for feature effect estimation is empirically negligible compared to the variance reduction gained from larger sample sizes. This provides a clear directive for XAI practitioners: use training data to get more stable and reliable feature effect estimates.
From the abstract
Global feature effects such as PD and ALE plots are widely used to interpret black-box models. However, they are only estimates of true underlying effects, and their reliability depends on multiple sources of error. Despite the popularity of global feature effects, these error sources are largely unexplored. In particular, the practically relevant question of whether to use training or holdout data to estimate feature effects remains unanswered. We address this gap by providing a systematic, est