predict_surrogate
function is added to provide easier interface of accessing lime/iml/localModel implementations of the LIME method. (this function wes moved toDALEXtra
)
- Fixes the bug in handling the case where all extracted variables are constant
- Changed default theme for plots, more consistent with other DrWhy packages
- Add
plot_interpretable_feature
function to plot discretizations. - Fixes bugs related to feature extraction.
- Changes the method of sampling new observations (features to permute are chosen from Beroulli distribution).
- Changes dependencies:
DALEX2
toDALEX
,ceterisParibus2
toingredients
.
- Vignettes updated. In particular, seed is now set before using randomForest.
- Plot function now shows a message and returns and empty plot if all effects are equal to 0.
- dev.ratio from glmnet is now returned as one of the columns of surrogate_model_explainer.
- Explanation models is now fitted to raw predictions, not difference between predictions and model mean.
- Fixed bug related to one-colum model matrix in glmnet.
- Bug fixed in example.
- Improved examples.
- Fixed issues with examples.
- Improved documentation.
- Fixed a bug related to interpretable numerical inputs (missing as.numeric)
- Added tests for 100% coverage.
- Set.seed before fitting cv.glmnet.
- New vignette describing methodology.
- Weighting option added.
- Improved documentation.
- Improvements to plots
- Multi-model plot option added.
stringr
dependency removed.
- Same interpretable features are used for every response level.
rpart
dependency replaced withpartykit
.
- Print method added.
- Updates link in README.
- Adds help for the package.
- Fixes bug related to number of levels when sampling new observation.
- Sets seed in tests.
- Minor updates to documentation (including paper/paper.md) and unit tests.
- Fixed bugs related to column names in case when response in is new_observation.
- Unit tests added.
- Vignettes added.
- Major changes: interpretable features are created using decision trees-based discretization.
- Now-unnecessary dependencies dropped.
- Plots are now drawn around intercept.
- Fixes bug in observation enconding.
- Improves and simplifies encoding.
- Improves plot.
- Improved plot labels.
- Added possibility of coding variables as (baseline / lower or higher) or ( baseline / lower / higher).
- Added option to skip loess smoothing.
- Removed option of setting
family
argument in glmnet.
- Added a
NEWS.md
file to track changes to the package.