Permutation Feature Importance Lab (2D)

A trained classifier's decision only matters as much as the features it actually uses — permutation importance finds out which ones those are without ever opening the model up. This simulator holds a batch of points in a 3-feature space where a fixed classification rule labels them, then repeatedly shuffles one feature column at a time across the whole dataset while leaving the true labels and the other two features untouched. Two of the features live on a pannable, zoomable 2D scatter plot; the third lives on its own animated 1D rail beneath it. Each shuffle is animated directly on the points: they slide to their new shuffled value on whichever panel represents the tested feature, and any point that becomes misclassified as a result turns red. Averaging the resulting accuracy drop over K independent shuffles per feature — with the spread reported as a standard deviation — produces the importance bars, exposing that this particular rule leans heavily on two of the three features and almost ignores the third.