Interpret

Added
- support for specifying outer bags
Changed
- exceptions raised in the joblib child processes will be re-raised in the main process rather than be expressed as a TerminatedWorkerError
- small additional improvements in memory compression
- small improvements in maximizing the benefit of the privacy budget for Differentially Private EBMs
Fixed
- fixed segfault that was occurring in the Anaconda build
- fixed a bug that would prevent Differentially Private EBMs from using the exclude parameter

Added
- support for visualizations in streamlit
Fixed
- fixed dangling pointer issue in call to CalcInteractionStrength

Added
- alternative objective functions: poisson_deviance, tweedie_deviance, gamma_deviance, pseudo_huber, rmse_log (log link)
- greediness __init__ parameter that allows selecting a behavior between cyclic boosting and greedy boosting
- smoothing_rounds __init__ parameter
- added type hints to the EBM __init__ parameters and class attributes
- init_score parameter to allow boosting and prediction on top of a previous model
- multiclass support in merge_ebms
- ability to monotonize features using post process model editing
Changed
- default BaseLinear regressor is changed from Lasso to LinearRegression class
- placed limits on the amount of memory used to find interactions with high cardinality categoricals
Fixed
- validation_size of 0 is now handled by disabling early_stopping and using the final model
Breaking Changes
- replaced the __init__ param "mains" with "exclude"
- removed the binning __init__ param as this functionality was already fully supported in feature_types
- removed the unused zero_val_count attribute and n_samples attribute
- renamed the noise_scale_ attribute to noise_scale_boosting_ and added noise_scale_binning_ to DPEBMs

Fixed
- fix the issue that the shared library would only work on newer linux versions

Added
- Mac m1 support in conda-forge
- SPOTGreedy prototype selection (PR 392)
Fixed
- fix visualization when both cloud and non-cloud environments are detected (PR 210)
- fix ShapTree bug where it was treating classifiers as regressors
- resolve scikit-learn warnings occurring when models were trained using Pandas DataFrames
- change the defaults to prefer 'continuous' over 'nominal' when a feature has 1 or 2 unique float64 values
Breaking Changes
- in the blackbox and greybox explainers, change from accepting a predict_fn to
accepting either a model or a predict_fn
- feature type 'categorical' has been renamed to 'nominal' for the remaining
feature_type parameters in the package (EBMs were already using 'nominal')
- removed the unused sampler parameters to the Explainer classes

Added
- Full Complexity EBMs with higher order interactions supported: GA3M, GA4M, GA5M, etc...
3-way and higher-level interactions lose exact global interpretability, but retain exact local explanations
Higher level interactions need to be explicitly specified. No automatic FAST detection yet
- Mac m1 support
- support for ordinals
- merge_ebms now supports merging models with interactions, including higher-level interactions
- added classic composition option during Differentially Private binning
- support for different kinds of feature importances (avg_weight, min_max)
- exposed interaction detection API (FAST algorithm)
- API to calculate and show the importances of groups of features and terms.
Changed
- memory efficiency: About 20x less memory is required during fitting
- predict time speed improvements. About 50x faster for Pandas CategoricalDType,
and varying levels of improvements for other data types
- handling of the differential privacy DPOther bin, and non-DP unknowns has been unified by having a universal unknown bin
- bin weights have been changed from per-feature to per-term and are now multi-dimensional
- improved scikit-learn compliance: We now conform to the scikit-learn 1.0 feature names API by using
self.feature_names_in_ for the X column names and self.n_features_in_.
We use the matching self.feature_types_in_ for feature types, and self.term_names_ for the additive term names.
Fixed
- merge_ebms now distributes bin weights proportionally according to volume when splitting bins
- DP-EBMs now use sample weights instead of bin counts, which preserves privacy budget
- improved scikit-learn compliance: The following __init__ attributes are no longer overwritten
during calls to fit: self.interactions, self.feature_names, self.feature_types
- better handling of floating point overflows when calculating gain and validation metrics
Breaking Changes
- EBMUtils.merge_models function has been renamed to merge_ebms
- renamed binning type 'quantile_humanized' to 'rounded_quantile'
- feature type 'categorical' has been specialized into separate 'nominal' and 'ordinal' types
- EBM models have changed public attributes:
-
feature_groups_ -> term_features_
global_selector -> n_samples_, unique_val_counts_, and zero_val_counts_
domain_size_ -> min_target_, max_target_
additive_terms_ -> term_scores_
bagged_models_ -> BaseCoreEBM has been depricated and the only useful attribute has been moved
into the main EBM class (bagged_models_.model_ -> bagged_scores_)
feature_importances_ -> has been changed into the function term_importances(), which can now also
generate different types of importances
preprocessor_ & pair_preprocessor_ -> attributes have been moved into the main EBM model class (details below)

- EBMPreprocessor attributes have been moved to the main EBM model class
-
col_names_ -> feature_names_in_
col_types_ -> feature_types_in_
col_min_ -> feature_bounds_
col_max_ -> feature_bounds_
col_bin_edges_ -> bins_
col_mapping_ -> bins_
hist_counts_ -> histogram_counts_
hist_edges_ -> histogram_edges_
col_bin_counts_ -> bin_weights_ (and is now a per-term tensor)