Introduced on day 17
Impurity importance
A tree model's built-in importance: the total split gain (impurity reduction) credited to a feature, computed on the training rows.
In detail
It is cheap and always available for trees, but it is biased toward continuous and many-valued features, counts gain on training rows (so overfit noise earns credit), and splits credit unpredictably among correlated features. Prefer permutation importance on held-out data for ranking (Strobl et al., 2007).