Classical ML, validation, metrics, interpretation, debugging · insight-0184

Report importance as a range across equally good models, not one model's ranking

Situation
Several models with nearly identical validation scores rank features differently, and you must present the drivers.
Action
Collect importances across near-optimal models (different seeds, families or hyperparameters within a tolerance of the best score) and report the range per feature.
Reason
Many models can fit the data equally well while relying on variables to very different degrees, so reliance is a property of the set, not the data.
Boundary
Driver reports, regulatory model documentation and fairness reviews.

Symptom or risk

Picking one model's ranking presents an arbitrary story as fact; the next retrain contradicts it.

When one model is mandated for deployment, its own importances describe what is deployed, but note the alternative explanations.

Ranges are honest but less tidy than a single bar chart.

Treating a stable ranking across seeds of one model family as robustness across all good models.

Model class reliance measures how much any well-performing model in a class may rely on a variable, motivated by the multiplicity of good models.
All Models are Wrong, but Many are Useful: Learning a Variable's Importance by Studying an Entire Class of Prediction Models Simultaneously