v0.2.3.2: pypi fix

v0.2.3.2

v0.2.3.visualize trees inside xgboost

Version 0.2.3

Breaking Changes

• RandomForestClassifierExplainer and RandomForestRegressionExplainer will be
  deprecated: can now simply use ClassifierExplainer or RegressionExplainer and the
  mixin class will automatically be loaded.

New Features

• Now also support for visualizing individual trees for XGBoost models!
  (XGBClassifier and XGBRegressor). The XGBExplainer mixin class will be
  automatically loaded and make decisiontree_df(), decision_path() and plot_trees()
  methods available, the dashboard Decision Trees tab and components now also work for
  XGBoost models.
• new parameter n_jobs for calculations that can be parallelized (e.g. permutation importances)
• contrib_df, plot_shap_contributions: can order by global shap feature
  importance with sort='importance' (as well as 'abs', 'high-to-low'
  'low-to-high')
• added actual outcome to plot_trees (for both RandomForest and XGB)

Improvements

• optimized code for calculating permutation importance, adding possibility to calculate in parallel
• shap dependence component: if no color col selected, output standard blue dots instead of ignoring update

Other Changes

• added selenium browser based integration tests for dashboards (also working with github actions)
• added tests for multiclass classsification, DecisionTree and ExtraTrees models
• added tests for XGBExplainers
• added proper docstrings to explainer_methods.py

v0.2.2: fix compatibility with shap 0.36

Bug Fixes

• fix for shap v0.36: import approximate_interactions from shap.utils instead of shap.common
• kernel shap bug fixed
• contrib_df bug with topx fixed

0.2.1

Version 2.1:

Breaking Changes

• Removed ExplainerHeader from ExplainerComponents
  • so also removed parameter header_mode from ExplainerComponent parameters
  • You can now instead syncronize pos labels across components with a PosLabelSelector
    and PosLabelConnector.
• In regression plots instead of boolean ratio=True/False,
  you now pass residuals={'difference', 'ratio', 'log-ratio'}
• decisiontree_df_summary renamed to decisiontree_summary_df (in line with contrib_summary_df)

New Features

• added check all shap values >-1 and <1 for model_output=probability
• added parameter pos_label to all components and ExplainerDashboard to set
  the initial pos label
• added parameter block_selector_callbacks to ExplainerDashboard to block
  the global pos label selector's callbacks. If you already have PosLabelSelectors
  in your layout, this prevents clashes.
• plot actual vs predicted now supported only logging x axis or only y axis
• residuals plots now support option residuals='log-ratio'
• residuals-vs-col plot now shows violin plot for categorical features
• added sorting option to contributions plot/graph: sort={'abs', 'high-to-low', 'low-to-high'}
• added final prediction to contributions plot

Bug Fixes

• Interaction connector bug fixed in detailed summary: click didn't work
• pos label was ignored in explainer.plot_pdp()
• Fixed some UX issues with interations components

Improvements

• All State['tabs', 'value'] condition have been taken out of callbacks. This
  used to fix some bugs with dash tabs, but seems it works even without, so
  also no need to insert dummy_tabs in ExplainerHeader.
• All ExplainerComponents now have their own pos label selector, meaning
  that they are now fully self-containted and independent. No global dash
  elements in component callbacks.
• You can define the layout of ExplainerComponents in a layout() method instead
  of _layout(). Should still define _register_callbacks() to define callbacks
  so that all subcomponents that have been registered will automatically
  get their callbacks registered as well.
• Added regression self.units to prediction summary, shap plots,
  contributions plots/table, pdp plot and trees plot.
• Clearer title for MEAN_ABS_SHAP importance and summary plots
• replace na_fill value in contributions table by "MISSING"
• add string idxs to shap and interactions summary and dependence plots,
  including the violing plots
• pdp plot for classification now showing percentages instead of fractions

Other Changes

• added hide_title parameter to all components with a title
• DecisionPathGraphComponent not available for RandomForestRegression models for now.
• In contributions graph base value now called 'population average' and colored yellow.

v0.2.0.2

Bug Fix:

• Title parameter was not correctly handed down to ExplainerTabsLayout and ExplainerPageLayout.

v0.2.0.1

Bug fix:
supress callback exception

v0.2

version 0.2:

Breaking Changes

• InlineExplainer api has been completely redefined
• JupyterExplainerDashboard, ExplainerTab and JupyterExplainerTab have been deprecated

New Features

• Major rewrite and refactor of the dashboard code, now modularized into ExplainerComponents
  and ExplainerComposites.
• ExplainerComponents can now be individually accessed through InlineExplainer
• All elements of components can now be switched on or off or be given an
  initial value.
• Makes it much, much easier to design own custom dashboards.
• ExplainerDashboard can be passed an arbitrary list of components to
  display as tabs.

Better docs:

• Added sections InlineExplainer, ExplainerTabs, ExplainerComponents,
  CustomDashboards and Deployment
• Added screenshots to documentation.

Bug Fixes

• fixes residuals y-pred instead of pred-y

Improvements

• Random Index Selector redesigned
• Prediction summary redesigned
• Tables now follow dbc.Table formatting
• All connections between components now happen through explicit connectors
• Layout of most components redesigned, with all elements made hideable

0.1.13

Bug Fixes

• Fixed bug with GradientBoostingClassifier where output format of shap.expected_value was not not properly accounted for.

Improvements

• Cleaned up standalone label selector code
• Added check for shap base values to be between between 0 and 1 for model_output=='probability'

0.1.12.5

typos

0.1.12.4

fixed InlineExplainer in all import of dashboards