Global Explanations

Explanation

Global explanations monitoring checks whether the features the model relies on most have shifted over time, because a sudden change in what drives the model can signal data drift, gaming, or a broken pipeline even when headline accuracy still looks fine. In fraud detection, for example, if the model abruptly stops weighting transaction velocity and starts leaning on a spurious feature, that is an early warning worth investigating. It is measured as Feature Importance Drift = 1 - Spearman Correlation (Current_Top10_Features, Baseline_Top10_Features); for fraud detection, alert if the correlation drops below 0.7. In plain terms, you rank the model's top-10 features now versus at baseline, compute the Spearman rank correlation between the two orderings, and subtract from one - so stable rankings give low drift and reordered rankings give high drift. To implement it, compute global feature importances (e.g. via a consistent attribution method) on a recurring schedule against a fixed baseline, store the top-10 rankings and the drift value as evidence, and chart it. The threshold is <0.3 drift, equivalently correlation >0.7; crossing it (drift at or above 0.3, correlation at or below 0.7) triggers an alert and an investigation into whether the shift reflects legitimate change or a problem requiring revalidation or retraining. Priority is Later. Aligns with RBI Understandable by Design, NIST MEASURE 3.1 and 4.1, and ISO Clauses 6.2.2 and 9.1.