release v0.23.0

• fix sphinx doc build failures in interpret-community
• sort imports using isort
• add more dependencies to docs build to fix warnings
• remove numba dependency pin
• add more flake8 extensions
• fix multiple save()/load() bug
• add flake8-breakpoint to avoid code checkin with active breakpoints

release v0.22.0

• version bump for shap 0.40.0 and interpret-core 0.2.7
• removed the pin on lightgbm package due to recent build break with new lightgbm release 3.3.1 and fixed the serialization logic to handle the new lightgbm model version
• upgraded tensorflow and xgboost test dependencies
• added support for keras scikit classifier and regressor to model wrapper
• fixed nightly build breaking due to new scikit-learn package which breaks model function serialization

release v0.21.0

• add an explanation adapter to integrate our explanations with other frameworks
• update interpret-community to interpret-core 0.2.6
• change explainers tabular explainer runs based on gpu flag
• fix model wrapper to handle a pytorch binary classification model that only outputs probabilities for positive class
• add test for old explanation dashboard and interpret dashboard
• fix nightly build breaking due to new scikit-learn package, which breaks model function serialization, by always serializing the model directly instead of the function if a model was passed to us

BREAKING CHANGE: function wrap_model now just returns the wrapped model

release v0.20.0

• Removed old ExplanationDashboard. The old namespace still exists but widget won't display anything and will print a warning now. Please use the ExplanationDashboard from the raiwidgets package instead.

  Please install raiwidgets from pypi by running:

  pip install --upgrade raiwidgets

  The dashboard can be run with the same parameters in the new namespace:

  from raiwidgets import ExplanationDashboard

  For more information on the new widget please see: https://github.com/microsoft/responsible-ai-widgets

• Fixed raw aggregated explanation failing to compute when transformations passed to mimic explainer and include_local=False

release v0.19.3

• emergency hotfix to pin numba to less than 0.54.0 to fix shap failures
• update to check correct version of RAPIDS
• update GPU SHAP for kmeans sampling code from cuML
• fix explanation dashboard failing to run on dataset with boolean target labels

release v0.19.2

• fix aggregation logic for raw global importance values to be consistent with local importance values

release v0.19.1

• fix error when creating a raw explanation from an engineered explanation that does not have a DatasetsMixin

release v0.19.0

• Update interpret-core to 0.2.5
• Add predicted and predicted probabilities to raw explanations to allow model performance tab and other tabs to show information related to y_pred/y_pred_proba for raw explanations
• Update cuML version for GPUKernelExplainer
• Remove shap DenseData from supported input data
• Use scipy logit() function in mimic explainer to fix divide by zero error

release v0.18.1

• includes fix to implement sparse case for methods:
  • get_ranked_local_values
  • get_ranked_local_names
  • get_local_importance_rank

and compress local importance values to dense format based on whether it will be more optimal storage when converting an engineered explanation to a raw explanation

• remove another spurious cuML warning message on library import.

release v0.18.0

• upgrade shap to 0.39.0, which deprecated python 3.5 and added python 3.8 support
• Add cuML GPU SHAP KernelExplainer
• update abstract classes to use ABC instead of ABCMeta
• remove warning on mimic explainer relating to categorical features
• Add test case of serializing pandas timestamp
• add sparse feature importance support for lightgbm surrogate model
• add support for drop parameter in one hot encoder