Variable Importance

Extracts from an "EffectData" object a simple variable importance measure, namely the (bin size weighted) variance of the partial dependence values, or of any other calculated statistic (e.g., "pred_mean" or "y_mean"). It can be used via update.EffectData(, sort_by = "pd") to sort the variables in decreasing importance. Note that this measure captures only the main effect strength. If the importance is calculated with respect to "pd", it is closely related to the suggestion of Greenwell et al. (2018).

effect_importance(x, by = NULL)

Arguments

x

Object of class "EffectData".

by

The statistic used to calculate the variance for. One of 'pd', 'pred_mean', 'y_mean', 'resid_mean', or 'ale' (if available). The default is NULL, which picks the first available statistic from above list.

Value

A named vector of importance values of the same length as x.

References

Greenwell, Brandon M., Bradley C. Boehmke, and Andrew J. McCarthy. 2018. A Simple and Effective Model-Based Variable Importance Measure. arXiv preprint. https://arxiv.org/abs/1805.04755.

See also

update.EffectData()

Examples

fit <- lm(Sepal.Length ~ ., data = iris)
M <- feature_effects(fit, v = colnames(iris)[-1], data = iris)
effect_importance(M)
#> Importance via weighted variance of 'pd'
#>  Sepal.Width Petal.Length  Petal.Width      Species 
#>   0.04579408   2.11662529   0.05714555   0.18456124