flashlight (version 1.0.0)

light_profile2d: 2D Partial Dependence and other 2D Profiles

Description

Calculates different types of 2D-profiles across two variables. By default, partial dependence profiles are calculated (see Friedman). Other options are response, predicted values, and residuals. The results are aggregated by (weighted) means.

Usage

light_profile2d(x, ...)
# S3 method for default
light_profile2d(x, ...)
# S3 method for flashlight
light_profile2d(
  x,
  v = NULL,
  data = NULL,
  by = x$by,
  type = c("partial dependence", "predicted", "response", "residual", "shap"),
  breaks = NULL,
  n_bins = 11L,
  cut_type = "equal",
  use_linkinv = TRUE,
  counts = TRUE,
  counts_weighted = FALSE,
  pd_evaluate_at = NULL,
  pd_grid = NULL,
  pd_indices = NULL,
  pd_n_max = 1000L,
  pd_seed = NULL,
  ...
)
# S3 method for multiflashlight
light_profile2d(
  x,
  v = NULL,
  data = NULL,
  type = c("partial dependence", "predicted", "response", "residual", "shap"),
  breaks = NULL,
  n_bins = 11L,
  cut_type = "equal",
  pd_evaluate_at = NULL,
  pd_grid = NULL,
  ...
)

Value

An object of class "light_profile2d" with the following elements:

• data A tibble containing results.

• by Names of group by variables.

• v The two variable names evaluated.

• type Same as input type. For information only.

Arguments

x

An object of class "flashlight" or "multiflashlight".

...

Further arguments passed to formatC() in forming the cut breaks of the v variables. Not relevant for partial dependence profiles.

v

A vector of exactly two variable names to be profiled.

data

An optional data.frame.

by

An optional vector of column names used to additionally group the results.

type

Type of the profile: Either "partial dependence", "predicted", "response", or "residual".

breaks

Named list of cut breaks specifying how to bin one or more numeric variables. Used to overwrite automatic binning via n_bins and cut_type. Ignored for non-numeric v.

n_bins

Approximate number of unique values to evaluate for numeric v. Can be an unnamed vector of length 2 to distinguish between v.

cut_type

Should numeric v be cut into "equal" or "quantile" bins? Can be an unnamed vector of length 2 to distinguish between v.

use_linkinv

Should retransformation function be applied? Default is TRUE.

counts

Should observation counts be added?

counts_weighted

If counts is TRUE: Should counts be weighted by the case weights? If TRUE, the sum of w is returned by group.

pd_evaluate_at

An named list of evaluation points for one or more variables. Only relevant for type = "partial dependence".

pd_grid

An evaluation data.frame with exactly two columns, e.g., generated by expand.grid(). Only used for type = "partial dependence". Offers maximal flexibility.

pd_indices

A vector of row numbers to consider in calculating partial dependence profiles. Only used for type = "partial dependence".

pd_n_max

Maximum number of ICE profiles to calculate (will be randomly picked from data). Only used for type = "partial dependence".

pd_seed

Integer random seed used to select ICE profiles. Only used for type = "partial dependence".

Methods (by class)

• light_profile2d(default): Default method not implemented yet.

• light_profile2d(flashlight): 2D profiles for flashlight.

• light_profile2d(multiflashlight): 2D profiles for multiflashlight.

Details

Different binning options are available, see arguments below. For high resolution partial dependence plots, it might be necessary to specify breaks, pd_evaluate_at or pd_grid in order to avoid empty parts in the plot. A high value of n_bins might not have the desired effect as it internally capped at the number of distinct values of a variable.

For partial dependence and prediction profiles, "model", "predict_function", "linkinv" and "data" are required. For response profiles it is "y", "linkinv" and "data". "data" can also be passed on the fly.

References

Friedman J. H. (2001). Greedy function approximation: A gradient boosting machine. The Annals of Statistics, 29:1189–1232.

See Also

light_profile(), plot.light_profile2d()

Examples

fit_part <- lm(Sepal.Length ~ Species + Petal.Length, data = iris)
fl_part <- flashlight(
  model = fit_part, label = "part", data = iris, y = "Sepal.Length"
)

# No effect of Petal.Width
plot(light_profile2d(fl_part, v = c("Petal.Length", "Petal.Width")))

# Second model includes Petal.Width
fit_full <- lm(Sepal.Length ~ ., data = iris)
fl_full <- flashlight(
  model = fit_full, label = "full", data = iris, y = "Sepal.Length"
)
fls <- multiflashlight(list(fl_part, fl_full))

plot(light_profile2d(fls, v = c("Petal.Length", "Petal.Width")))