light_profile2d
for 2D plots of partial dependence and other profiles (residuals, predictions, response, SHAP).Added logo
All zzz_names
arguments are deprecated and are set as options(flashlight.zzz_names = ...)
instead.
light_ice
and light_scatter
now use override.aes
to suppress alpha
in legend.
Greatly improved error messages.
Bug fixes.
Another maintainance release to remove dependency on ggpubr.
This is a maintainance release to deal with an issue with XGBoost.
This is a documentation and maintainance release.
Switched to dplyr 1.0.0.
Added 270+ unit tests.
Added vignettes on how to use flashlight with caret
and mlr3
.
light_breakdown
have been slightly improved.light_effects
and light_profile
did not work correctly if pred
are provided together with a pd_indices
.
Using pred
in light_effects
and light_profile
with a multiflashlight will now produce an informative error.
light_interaction
now uses a default grid size of 200 instead of 30 to calculate partial dependency profiles. Furthermore, the used sample size to calculate the interaction statistics have been increased from 300 to 1000. The two changes improve considerably the stability of the results at the price of a substantial increase in runtime. For slow prediction functions, choose lower values as suitable.Like for light_effects
, the plot method of light_profile
has received an argument show_points
to suppress the plotting of points.
The plot method for light_performance
has received a geom
argument. It allows to switch from bars to points.
Now requires at least tidyr >= 1.0.0 since expand_grid
is used to calculate ICE profiles.
R dependency has been lowered from 3.5 to 3.1.
Removed suggested packages caret and lubridate.
Emergency release. Fixes a problem introduced with ggplot 3.3.0 when plotting light_importance and light_interaction objects.
light_global_surrogate
and plot.light_global_surrogate
to create a global surrogate tree model that explains predictions from the original model in a simple to understand way.Added light_scatter
and plot.light_scatter
to create scatter plots of response, predictions, residuals, or SHAP values by some variable.
Static (approximate) SHAP values can be added to a (multi-)flashlight by calling add_shap
.
light_importance
, light_profile
, and light_scatter
are now able to work with SHAP values.
The plot functions of light_importance
, light_interaction
, and light_profile
now prints informative labels to show what is actually calculated.
Added argument “show_points” to plot.light_effects
to control if points should be added to the lines or not. Default is TRUE
.
Removed currently unused output columns “value_shuffled” and “value_before” of light_importance
due to its incompatibility with the new “shap” importances.
Added argument “type_name” to light_interaction
. Its output now also contains a slot type
as well as type_name
.
Added Friedman’s H statistic (global and pairwise interaction strength) to light_interaction
with variants.
Added random permutation visit strategy to light_breakdown
, serving as approximate SHAP.
Added more options how to center light_ice
: Mean center each profile to the same value (within “by” groups) and 0-mean center each profile.
Added option rotate_x
to plot_light_breakdown
and plot_light_importance
.
Added function grouped_center
to allow grouped and weighted 0-mean centering of a numeric variable.
Added function grouped_count
to allow grouped and weighted counts.
Added function grouped_weighted_mean
for fast grouped weighted means.
response
, residuals
, and predict
now have a method for multiflashlights.
Combined arguments center = TRUE/FALSE
and center_at
to one argument center
with default “no”. This affects light_ice
, and light_profile
.
order_by_importance
argument to light_breakdown
has been replaced by visit_strategy
.
Removed top_m
argument of light_breakdown
as it does not make sense.
auto_cut
, the workhorse behind quantile binning a numeric input variable x, is now using ecdf based quantiles in order to ensure all evaluation points are in the domain of x.
Centering at “first”, “middle”, and “last” in light_ice
now anchors the curves at 0 to be in line with other implementations.
light_ice
was based on dplyr::crossing
, unintentionally throwing away duplicate reference rows. This is now replaced by dplyr::expand_grid
.Added light_interaction
, a measure of overall interaction strength per covariable based on standard deviation across c-ICE curves.
Added the option m_repetitions = 1
to light_importance
. Set to higher value to get more stable estimates of variable importance as well as standard errors for small data.
Added the option digits
to plot_counts
with default 0. Helps to format large counts or fractional counts.
Added to option center_at
to light_ice
in order to control if ICE curves should be centered at position “first”, “middle”, or “last”. Similarly, added pd_center_at
to light_profile
.
plot
-method of light_ice
.Added type = "ale"
to light_profile
to show accumulated local effects plots (ALE plots). They are now also calculated by light_effects
. In the plot
method, set use = "all"
to show all profiles. The default is all without ALE plots.
Added the possibility to center ICE curves at first evaluation point to improve visibility of interactions. Together with partial dependence with quartile aggregation, this is an interesting possibility to detect interactions.
grouped_stats
has received an argument value_name
.
Option cut_type
was not active for partial dependence profiles.
Option cut_tpye
was not active for the multiflashlight method of light_profile
.
zero_counts
argument in plot_counts
.zero_counts
in plot_counts
and plot.light_effects
had no effect for single flashlights with no “by” variable. This is fixed.This is the initial release.