mboost (version 2.9-11)

varimp: Variable Importance

Description

In-bag risk reduction per base-learner as variable importance for boosting.

Usage

# S3 method for mboost
varimp(object, ...)
# S3 method for varimp
plot(x, percent = TRUE, type = c("variable", "blearner"), 
  blorder = c("importance", "alphabetical", "rev_alphabetical", "formula"),
  nbars = 10L, maxchar = 20L, xlab = NULL, ylab = NULL, xlim, auto.key, ...)
# S3 method for varimp
as.data.frame(x, row.names = NULL, optional = FALSE, ...)

Value

An object of class varimp with available plot and as.data.frame methods.

Converting a varimp object results in a data.frame containing the risk reductions, selection frequencies and the corresponding base-learner and variable names as ordered factors (ordered according to their particular importance).

Arguments

object

an object of class mboost.

x

an object of class varimp.

percent

logical, indicating whether variable importance should be specified in percent.

type

a character string specifying whether to draw bars for variables ("variable", default) or base-learners ("blearner") in the model (no effect for a glmboost object).

blorder

a character string specifying the order of the base-learners in the plot. The default "importance" corresponds to the order of the base-learner importance, "alphabetical" and "rev_alphabetical" to alphabetical order, respectively its reverse, and "formula" to their order in the model formula.

nbars

integer, maximum number of bars to be plotted. If nbars is exceeded, least important variables / base-learners are summarized as "other".

maxchar

integer, maximum number of characters in bar labels.

xlab

text for the x-axis label. If not set (default is NULL) x-axis label is generated automatically depending on argument percent.

ylab

text for the y-axis label. If not set (default is NULL) y-axis label is generated automatically depending on argument type.

xlim

the x limits of the plot. Defaults are from 0 to total reduction, or from 0 to 1 for percent = TRUE. (In case of negative risk reductions, default limits are from total negative to total positve reduction, or the latter normalized by the total absolute reduction for percent = TRUE.)

auto.key

logical, or a list passed to lattice::barchart. By default auto.key=TRUE provides automatically generated legends showing the underlying base-learners in the stacked barchart (type = "variable"). If there is an unique base-learner for each variable(-interaction), auto.key = FALSE is default setting. For type = "blearner" the argument has no effect at all.

...

additional arguments passed to lattice::barchart.

row.names

NULL or a character vector giving the row names for the data frame. Missing values are not allowed.

optional

logical. If TRUE, setting row names and converting column names (to syntactic names: see make.names) is optional.

Author

Tobias Kuehn (), Almond Stoecker ()

Details

This function extracts the in-bag risk reductions per boosting step of a fitted mboost model and accumulates it individually for each base-learner contained in the model. This quantifies the individual contribution to risk reduction of each base-learner and can thus be used to compare the importance of different base-learners or variables in the model. Starting from offset only, in each boosting step risk reduction is computed as the difference between in-bag risk of the current and the previous model and is accounted for the base-learner selected in the particular step.

The results can be plotted in a bar plot either for the base-learners, or the variables contained in the model. The bars are ordered according to variable importance. If their number exceeds nbars the least important are summarized as "other". If bars are plotted per variable, all base-learners containing the same variable will be accumulated in a stacked bar. This is of use for models including for example seperate base-learners for the linear and non-linear part of a covariate effect (see ?bbs option center=TRUE). However, variable interactions are treated as individual variables, as their desired handling might depend on context.

As a comparison the selection frequencies are added to the respective base-learner labels in the plot (rounded to three digits). For stacked bars they are ordered accordingly.

Examples

data(iris)
### glmboost with multiple variables and intercept
iris$setosa <- factor(iris$Species == "setosa")
iris_glm <- glmboost(setosa ~ 1 + Sepal.Width + Sepal.Length + Petal.Width +
                         Petal.Length,
                     data = iris, control = boost_control(mstop = 50), 
                     family = Binomial(link = c("logit")))
varimp(iris_glm)
### importance plot with four bars only
plot(varimp(iris_glm), nbars = 4)

### gamboost with multiple variables
iris_gam <- gamboost(Sepal.Width ~ 
                         bols(Sepal.Length, by = setosa) +
                         bbs(Sepal.Length, by = setosa, center = TRUE) + 
                         bols(Petal.Width) +
                         bbs(Petal.Width, center = TRUE) + 
                         bols(Petal.Length) +
                         bbs(Petal.Length, center = TRUE),
                     data = iris)
varimp(iris_gam)
### stacked importance plot with base-learners in rev. alphabetical order
plot(varimp(iris_gam), blorder = "rev_alphabetical")

### similar ggplot
if (FALSE) { 
library(ggplot2)
ggplot(data.frame(varimp(iris_gam)), aes(variable, reduction, fill = blearner)) + 
    geom_bar(stat = "identity") + coord_flip() }