breaking change: calculate_variable_splits() now treatsinteger variables as categorical. This change is propagated to ceteris_paribus(),partial_dependence(),accumulated_dependence(),conditional_dependence(),aggregate_profiles(),DALEX::predict_profile(),DALEX::model_profile()
fix an error in ceteris_paribus /calculate_variable_splits when tidymodels usesinteger variables #145
fix an error in show_observations #148 . This change is propagated to DALEX::plot.predict_profile() #540 .
fix #149 by replacing all class(x) = "y" withis(x, "y")
added facet_scales parameter toplot.aggregated_profiles_explainer ('free_x'by default) #138 and plot.ceteris_paribus_explainer ('free_x'or 'free_y' by default, depending on plot type) #136
fixes explanations when data has one column #137
code and documentation maintenance #130
fixed an error when N = NULL inpartial_dependence() etc. #134
plot.ceteris_paribus_explainer now by default for categorical variables plots profiles (not lines -prev default- nor bars)ALE plots are now centered around average y_hat #126
colors from DrWhy color palette is used for CP #125
default subtitle value in plot.fi changed to NULL from NA (unification)
now in the ceteris_paribus function one can specify how grid points shall be calculated, seevariable_splits_type
ceteris_paribus and aggregates are now working with missing data, this solves #120 plot(ceteris_paribus) change default colorto label or ids if more than one profile is detected, this solves #123 ceteris_paribus has now argumentvariable_splits_with_obs which included values fromnew_observations in the variable_splits, this solves #124
deprecate n_sample argument infeature_importance (now it’s N) #113
plot_profile now handles multilabel models
DALEX is moved to Suggests as in #112 plot_categorical_ceteris_paribus can plot bars (again)add bind_plots function
support R v4.0 and depend on R v3.5 to comply with DALEX
new arguments title and subtitle in several plots
change dependency to dependence #103
ceteris_paribus profiles are now working for categorical variablesshow_profiles, show_observations,show_residuals are now working for categorical variables
synchronisation with changes in DALEX 0.5
new argument desc_sorting inplot.variable_importance_explainer #94
feature_importance now does 15permutations on each variable by default. Use the Bargument to change this numberadded boxplots to plot.feature_importance andplotD3.feature_importance that showcase the permutation data
in aggregate_profiles: preserve _x_ column factor order and sort its values #82
aggregate_profiles use now gaussian kernel smoothing. Use the span argument for fine control over this parameter (#79 )change variable_type and variablesarguments usage in the aggregate_profiles,plot.ceteris_paribus andplotD3.ceteris_paribus
remove variable_type argument fromplotD3.aggregated_profiles (now the same as inplot.aggregated_profiles)
Kasia Pekala is moved as contributor to the DALEXtra asaspect_importance is moved to DALEXtra as well (See v0.3.12 changelog )
added Travis-CI for OSX
fixed rounding problem in the describe function (#76 )
aspect_importance is moved to DALEXtra (#66 )examples are updated in order to reflect changes intitanic_imputed from DALEX (#65 )
modified plot.aspect_importance - it can plot more than single figure
modified triplot, plot.aspect_importanceand plot_group_variables to add more clarity in plots and allow some parameterization
added triplot function that illustrates hierarchicalaspect_importance() groupings
changes in aspect_importance() functions
added back the vigniette for aspect_importance()
change only_numerical parameter tovariable_type in functions aggregated_profiles(), cluster_profiles(), plot() and others, as requested in #15
Natural language description generated with describe()function for ceteris_paribus(),feature_importance() and aggregate_profiles()explanations.
aggregated_profiles_conditional andaggregated_profiles_accumulated are rewritten with some code fixes
a new version of lime is implemented in thelime()/aspect_importance() function.
Kasia Pekala and Huber Baniecki are added as contributors.
new feature #29 . Feature importance now takes an argument B that replicates permutations B times and calculates average from drop loss.
plotD3 now supports Ceteris Paribus Profiles.feature_importance now can takevariable_grouping argument that assess importance of group of featuresfix in ceteris_paribus, now it handles models with just one variable
fix #27 for multiple rows
show_profiles and show_residuals functions extend Ceteris Paribus Plots.show_aggreagated_profiles is renamed toshow_aggregated_profilescentering of ggplot2 title
added new functions describe() andprint.ceteris_paribus_descriptions() for text based descriptions of Ceteris Paribus explainers
plot.ceteris_paribus_explainer works now also for categorical variables. Use the only_numerical = FALSE to force bars
added references to PM VEE
partial_profiles(), accumulated_profiles()and conditional_profiles for variable effectsmajor changes in function names and file names
ceteris_paribus_2d extends classical ceteris paribus profilesceteris_paribus_oscillations calculates oscilations for ceteris paribus profilesfixed examples and file names
cluster_profiles helps to identify interactionspartial_dependency calculates partial dependency plotsaggregate_profiles calculates partial dependency plots and much more
port of model_feature_importance andmodel_feature_response from DALEX toingredients
added tests