Compute ICE-based variable importance scores for the predictors in a model. (This function is meant for internal use only.)
vi_ice(object, ...)# S3 method for default
vi_ice(object, feature_names, FUN = NULL,
var_fun = NULL, ...)
A fitted model object (e.g., a "randomForest" object).
Additional optional arguments to be passed onto
partial.
Character string giving the names of the predictor variables (i.e., features) of interest.
Deprecated. Use var_fun instead.
List with two components, "cat" and "con",
containing the functions to use to quantify the variability of the feature
effects (e.g., partial dependence values) for categorical and continuous
features, respectively. If NULL, the standard deviation is used for
continuous features. For categorical features, the range statistic is used
(i.e., (max - min) / 4). Only used when method = "pdp" or
method = "ice".
A tidy data frame (i.e., a "tibble" object) with two columns,
Variable and Importance, containing the variable name and its
associated importance score, respectively.
Similar to the PDP-based method, this approach to computing VI scores is based on quantifying the relative "flatness" of the individual conditional expectation (ICE) curves of each feature and then averaging them together. It is model-agnostic and can be applied to any supervised learning algorithm. By default, relative "flatness" is defined by computing the standard deviation of the y-axis values for each PDP for numeric features; for categorical features, the default is to use range divided by 4. This can be changed via the `var_fun` argument. See Greenwell et al. (2018) for details and additional examples.
#' @references Greenwell, B. M., Boehmke, B. C., and McCarthy, A. J. A Simple and Effective Model-Based Variable Importance Measure. arXiv preprint arXiv:1805.04755 (2018).