Plot feature importance across multiple models

Each model's importance is rescaled on its own: its largest value becomes 100, or, when no value is positive (a forest's permutation importance can be negative throughout), its largest magnitude becomes -100. A factor predictor appears once, under its own name, for every model: the largest of its design columns stands for it where a method ranks those columns separately (ridge, lasso, elastic net and xgboost). A predictor a model was given but did not use scores zero for that model; one it was never given has no bar for it. Tree, forest and boost models are given the variables of an interaction such as wt:hp rather than the interaction itself, so it has no bar for them. Features are ranked on their mean importance over the models that were given them.

Usage

tl_plot_importance_comparison(..., top_n = 10, names = NULL)

Arguments

...

tidylearn model objects to compare

top_n

Number of top features to display (default: 10)

names

Optional character vector of model names, one unique name per model

Value

A ggplot object.

Examples

# \donttest{
m1 <- tl_model(iris, Sepal.Length ~ ., method = "forest")
m2 <- tl_model(iris, Sepal.Length ~ ., method = "boost")
tl_plot_importance_comparison(m1, m2, names = c("Forest", "Boost"))

# }