Release Notes¶
Version 0.17.3 (2023-02-12)¶
Improvements:
Implement
sample_weight
check in Scorecard class (Issue 228).
Bugfixes:
Fix
metric_missing
ignored in Scorecard class (Issue 226).
Dependencies:
Update RoPWR required version.
Version 0.17.2 (2022-12-15)¶
Improvements:
Modify max-pvalue and min_diff constraints for CP and MIP formulation to avoid suboptimal solutions.
Bugfixes:
Dependencies:
Update scikit-learn and ortools required versions.
Version 0.17.1 (2022-11-20)¶
New features:
Add parameter
cat_unknown
to assign values to the unobserved categories during training.
Improvements:
Add method
decision_function
toScorecard
(Issue 198).
Version 0.17.0 (2022-11-06)¶
New features:
Bugfixes:
Fix
ContinuousOptimalBinning
prebinning step when no prebinning splits were generated (Issue 205).
Version 0.16.1 (2022-10-28)¶
New features:
Outlier detector
YQuantileDetector
for continuous target (Issue 203).
Improvements:
Add support to solver SCS and HIGHS for optimal piecewise binning classes.
Unit testing outlier detector methods.
Bugfixes:
Pass
lb
andub
as keyword arguments to RoPWR fit method (required since ropwr>=0.4.0).
Version 0.16.0 (2022-10-24)¶
New features:
Treatment of special codes separately for optimal piecewise binning classes (Issue 191).
Improvements:
Allow plot
style="actual"
for stochastic optimal binning.Unit testing optimal piecewise binning classes (Issue 93).
Unit testing add macOS Monterey 12.
Bugfixes:
Version 0.15.1 (2022-09-06)¶
New features:
New parameter
show_bin_labels
for binning tables (Issue 180).
Version 0.15.0 (2022-07-20)¶
New features:
Optimal binning 2D support to categorical variables for binary and continuous target.
Improvements:
Bugfixes:
Fix
Scorecard.score
method when there are special and missing bins. (Issue 179).Fix x and y axis labels in
OptimalBinning2D
plots, x and y were interchanged.
Version 0.14.0 (2022-04-10)¶
New features:
Optimal binning 2D with continuous target.
Improvements:
Set tdigest and pympler dependencies as optional. This change avoids accumulation-tree issues faced by several users. Remove dill dependency.
New continuous binning objective function leading to improvements in regression metrics.
Bugfixes:
Fix binning 2D minimum difference constraints.
Tutorials:
Tutorial: optimal binning 2D with continuous target
Version 0.13.0 (2021-11-24)¶
New features:
Treatment of special codes separately for optbinning classes (Issue 115).
Bugfixes:
Various bug fixes for the
OptimalBinning2D
class. See Issue 138, for instance.
Tutorials:
Tutorial: optimal binning 2D with binary target
Version 0.12.2 (2021-10-03)¶
Improvements:
Do not store optimization solver instance as class attribute.
Do not store logger as a class attribute.
Version 0.12.0 (2021-08-28)¶
New features:
Optimal binning 2D with binary target.
Improvements:
Update bin string format in binning tables.
Simplify logic when
style="actual"
in binning table plots.
API changes:
Scorecard fit method arguments changed to the usual
(X, y)
: Issue 111
Version 0.11.0 (2021-05-28)¶
New features:
Counterfactual explanations for scorecard modelling.
Improvements:
Replace pickle by dill in save and load methods.
Bugfixes:
Parallel binning uses joblib: Issue 103
Fix custom
metric_special
andmetric_missing
in binning_transform_params.
Version 0.10.0 (2021-04-27)¶
New features:
Batch and streaming binning process.
Improvements:
Improve LocalSolver formulation for optimal binning with a binary target.
Bugfixes:
Version 0.9.2 (2021-03-12)¶
New features:
Binning process can update binned variables with new optimal binning object using method
update_binned_variable
.
Improvements:
Prevent large divisions to avoid overflow issues with int32 during Gini calculation.
Tutorials:
Tutorial: FICO Explainable Machine Learning Challenge - updating binning
Version 0.9.1 (2021-02-14)¶
New features:
Binning process can be constructed using OptimalBinning objects previously fitted. Method
fit_from_dict
.Binning process can process large datasets directly on disk. Allowed file formats are csv and parquet. Methods
fit_disk
,fit_transform_disk
andtransform_disk
.
Bugfixes:
Fix saving all OptBinning classes: Issue 77
Version 0.9.0 (2021-01-14)¶
New features:
Optimal piecewise polynomial binning.
New plotting option for binning table for binary and continuous target. Parameter
style
allows to represent the binning plot with the actual scale, i.e., actual bin widths.
Improvements:
Improve computation of p-values and binning table analysis for
ContinuousOptimalBinning
.
Tutorials:
Tutorial: optimal piecewise binning with binary target
Tutorial: optimal piecewise binning with continuous target
Bugfixes:
Fix sample weights bug: Issue 64
Version 0.8.0 (2020-09-18)¶
New features:
Scorecard monitoring supporting binning and continuous target.
OptimalBinning computes the Kolmogorov-Smirnov statistic.
Optimal binning classes show optimal monotonic trend information in the binning table analysis method.
ContinuousBinningTable adds method
analysis
.Scorecard incorporates methods
load
andsave
to serialize and deserialize a scorecard using pickle module.BinningProcess class supports multiprocessing via parameter
n_jobs
.
Tutorials:
Tutorial: Scorecard monitoring
Version 0.7.0 (2020-07-19)¶
New features:
Batch and streaming optimal binning.
New parameter
divergence
to select the divergence measure to maximize.
Tutorials:
Tutorial: optimal binning sketch with binary target
Tutorial: optimal binning sketch with binary target using PySpark
Bugfixes:
Catch error from Qhull library used by scipy.spatial.ConvexHull.
Version 0.6.1 (2020-06-07)¶
New features:
Options
add_special
andadd_missing
in all binning table plots.Prebinning methods’ parameters are accessible via
**prebinning_kwargs
.Add support MDLP algorithm for binary target.
Bugfixes:
Fix bug in solution when the status is not feasible or optimal for LocalSolver,
solver="ls"
.Fix several bugs for categorical variables with
user_splits
anduser_splits_fixed
.Fix bug in binning process when passing
user_splits
anduser_splits_fixed
via parameterbinning_fit_params
.
Version 0.6.0 (2020-05-24)¶
New features:
Scorecard development supporting binary and continuous target.
Plotting functions:
plot_auc_roc
,plot_cap
andplot_ks
.Optimal binning classes introduce
sample_weight
parameter in methodsfit
andfit_transform
.Optimal binning classes introduce two options for parameter
metric
in methodsfit_transform
andtransform
:metric="bins"
andmetric="indices"
.
Tutorials:
Tutorial: optimal binning with binary target - large scale.
Tutorial: Scorecard with binary target.
Tutorial: Scorecard with continuous target.
Version 0.5.0 (2020-04-13)¶
New features:
Scenario-based stochastic optimal binning.
New parameter
user_split_fixed
to force user-defined split points.
Tutorials:
Tutorial: Telco customer churn.
Tutorial: optimal binning with binary target under uncertainty.
Bugfixes:
Fix monotonic trend for non-auto mode in
MulticlassOptimalBinning
.
Version 0.4.0 (2020-03-22)¶
New features:
New
monotonic_trend
auto modes options: “auto_heuristic” and “auto_asc_desc”.New
monotonic_trend
options: “peak_heuristic” and “valley_heuristic”. These options produce a remarkable speedup for large size instances.Minimum Description Length Principle (MDLP) discretization algorithm.
Improvements:
BinningProcess
now supportspandas.DataFrame
as input X.New unit test added.
Version 0.3.0 (2020-03-13)¶
New features:
Class
OptBinning
introduces a new constraint to reduce dominating bins, using parametergamma
.Metrics HHI, HHI regularized and Cramer’s V added to
binning_table.analysis
method. Updated quality score.Added column min/max target and zeros count to
ContinuousOptimalBinning
binning table.Binning algorithms support univariate outlier detection methods.
Tutorials:
Tutorial: optimal binning with binary target. New section: Reduction of dominating bins.
Enhance binning process tutorials.
Version 0.2.0 (2020-02-02)¶
New features:
Binning process to support optimal binning of all variables in dataset.
Added
print_output
option tobinning_table.analysis
method.
Improvements:
New unit tests added.
Tutorials:
Tutorial: Binning process with Scikit-learn pipelines.
Tutorial: FICO Explainable Machine Learning Challenge using binning process.
Bugfixes:
Fix
OptBinning.information
print level default option.Avoid numpy.digitize if no splits.
Compute Gini in
binning_table.build
method.
Version 0.1.1 (2020-01-24)¶
Bugfixes:
Fix a bug in
OptimalBinning.fit_transform
when callingtranform
internally.Replace np.int by np.int64 in
model_data.py
functions to guarantee 64-bit integer on Windows.Fix a bug in
_chech_metric_special_missing
.
Version 0.1.0 (2020-01-22)¶
First release of OptBinning.