Model‐Agnostic Influential Outlier Metric

The influential outlier metric (IOM) provides model‐agnostic influential outlier detection. We define influence of an observation using a combination of SHapley Additive exPlanation (SHAP) values and the residual. Both are transformed using normalizing flows, changing their respective measures to Gaussian distributions. The transformed SHAP values and residuals define the IOM as a product of two independent random variables, allowing critical values for a given significance level . We illustrate the advantages and limitations of this approach in several models including the linear model, neural networks, the random forest and gradient‐boosted trees.