Machine learning models that explain their own predictions often produce a handful of conflicting explanations for the same call, and nothing in most systems tells you which one to believe.
A new paper, revised this month, builds a framework for ranking competing explanations instead of leaving the choice to chance, sorting each candidate by whether it raises or lowers prediction uncertainty, which direction it pushes the prediction, and where it falls relative to a decision boundary. The framework then applies eligibility rules, an optional two-way Pareto filter, and a ranking policy to narrow the field. The authors test it on Calibrated Explanations across classification and regression tasks, plus a second explanation generator called delta-CLUE, run against 41 benchmark datasets. On average, a prediction generates 11.57 to 21.75 candidate single-feature explanations, and 29.48 to 69.53 once combinations of features are allowed.
That candidate count is the real story: a diagnostic tool that hands a doctor dozens of technically valid but differently-flavored explanations for the same result isn't more transparent, it's just more confusing, unless something decides which explanation fits the question being asked. The paper's own test case, a hypothetical prostate-cancer prediction, shows two reasonable selection policies, one weighing confidence and uncertainty equally, one weighing confidence alone, disagreeing on which explanation to surface 28.7% of the time, even though they usually agree on the overall direction of the prediction.
This isn't a new explanation method, it's a sorting mechanism bolted onto existing ones, which is a modest contribution, but modest is fine: right now most explainable-AI tools just dump every candidate on the user and call it interpretability.