Variable importance

Variable importance is a quantitative measure of how much a given feature contributes to the predictive accuracy of a model. It is not merely a correlation coefficient; it captures the marginal loss in model performance when the feature is degraded or removed, revealing which inputs the model actually relies on rather than which inputs are merely statistically associated with the target. In random forests, variable importance is computed by measuring the increase in prediction error when the values of a feature are randomly permuted across the out-of-bag samples — a method that exposes the model's genuine dependence on that feature rather than its coincidental presence in the training data.

The concept is deceptively simple and dangerously seductive. Variable importance scores are often interpreted as causal indicators, but they are not: a feature can be important because it is a proxy for a truly causal variable, or because it is collinear with another feature that the model happens to select. The ranking of importance is sensitive to the correlation structure of the data, and different models may assign radically different importance scores to the same features. The measure is therefore a diagnostic tool, not a certificate of causal relevance. The most sophisticated applications of variable importance use it not to claim understanding but to identify where the model's reasoning is fragile and where human intervention is required.

The deeper question is whether any purely quantitative measure can capture what humans mean by 'importance.' A variable may be numerically marginal but conceptually central, or numerically dominant but epistemically trivial. The tension between statistical importance and explanatory importance is one of the persistent fractures in the epistemology of machine learning.