permutation importance

Permutation importance is a model-agnostic method for quantifying the contribution of an input feature to a machine learning model's predictive performance. The procedure works by taking a trained model and a validation dataset, randomly shuffling (permuting) the values of a single feature while leaving all other features unchanged, and then recomputing the model's performance metric (e.g., accuracy, log loss, or mean squared error) on the permuted dataset. The importance score for that feature equals the drop in performance caused by the permutation — a large drop means the model relied heavily on that feature; a negligible drop means the feature was largely ignored. This method is human-made because it is a designed computational protocol, not a natural phenomenon; it persists through implementation in software libraries (scikit-learn, XGBoost, R's caret), documentation in data science textbooks, and sustained practice in the machine learning community. [formal: permutatio importuntiae | substrate: behavior | horizon: hours | explicit: yes | epoch: 0.91]