Analyzing Error Sources in Global Feature Effect Estimation

Training data can be practically reliable for feature effect estimation: While holdout (validation) data is theoretically cleaner, the paper finds that biases from using training data are "empirically negligible and dominated by the impact of the usually higher sample size," suggesting practitioners don't need to reserve data solely for this interpretability task.

Sample size matters more than data source for reducing uncertainty: The research reveals that estimation variance (uncertainty in the estimates) is primarily driven by sample size rather than whether training or validation data is used, meaning larger datasets lead to more reliable feature effect plots regardless of data selection strategy.

Accumulated Local Effects (ALE) plots are more sensitive to sample size than Partial Dependence (PD) plots: The study finds that ALE estimates are "particularly sensitive" to sample size, which has practical implications for when these visualization methods will produce reliable interpretations of model behavior.

Interactions in data increase estimation uncertainty: The presence of feature interactions affects how much uncertainty appears in feature effect estimates, a factor that practitioners should consider when interpreting these plots, particularly for complex models.

Cross-validation offers a promising middle ground: The paper identifies cross-validation-based estimation as "a promising" approach, suggesting it may balance the theoretical advantages of holdout data with the practical benefits of using more data for estimation.