Releases: microsoft/responsible-ai-toolbox
Releases · microsoft/responsible-ai-toolbox
release raiwidgets 0.13.0
v0.13.0
- new features
-
Model Analysis
- add "Save amd switch" button to cohort creation
- pass classes parameter to both error analysis dashboard and error analysis manager in model analysis
- upgrade interpret-community to 0.20.0
-
Error Analysis
- update erroranalysis to 0.1.24
-
- bug fixes
-
Model Analysis
- fix issue with widgets rendering with 8px height
-
Error Analysis
- fix precision and recall score failing on string labels in error analysis
- fix multiclass precision and recall calculation for NA zero values in heatmap
- edit string in error analysis cohort creation
- remove text in heatmap suggesting to select features when in static view
- minor doc fix for examples in error analysis dashboard
-
Counterfactual
- fix horizontal scroll bar disappears for small screen in what-if panel
-
Fairness
- fix data issue caused robot page exception on fairnessDetails
-
- other
- fix broken link to getting started notebook
- fix broken image link on main readme
- fix the issues with sample notebooks markdown images
- delete old notebooks
- change tests to run on new notebooks, add new notebook to tests
release raiwidgets 0.12.1
v0.12.1
- new features
-
Error Analysis
- add disabled tree view and heatmap for static view
-
- bug fixes
-
Error Analysis
- fix class length causing labels to not be shown in simple view of Error Analysis
- fix cohort filter tooltip behind cohort info panel
-
Causal
- fix trace0 appearing on hover tooltip in plots
-
Counterfactual
- handle cases where counterfactuals failed to compute
- add tooltip for counterfactual info
-
- other
- some grammatical errors were corrected on documentation page
release raiwidgets 0.12.0
v0.12.0
- new features
-
Model Analysis
- add global cohort statistics
-
Error Analysis
- add metric selector dropdown
- add tree view parameters as sliders to features selector panel
- add tooltip to heatmap categories and left-justify the heatmap
-
Counterfactual
- upgrade dice-ml to 0.7.2
-
- bug fixes
-
Model Analysis
- fix cohort information in heatmap resizing in model analysis
- log warning when regression model has predict_proba()
-
Error Analysis
- fix error rate gradient for classification scenario
- fix quantile binning toggle should stay on when making actions
- fix multiple issues in instance view, including:
- when selecting multiple rows in correct and incorrect panels, and going to all selected, if a point is removed in the middle all of the points below get removed
- when removing points in the all selected panel, the count does not get updates in the correct/incorrect panels until they are selected
- when using what if, make points not selectable
- fix precision and recall calculation in the UI changing with number of bins
- decrease the lower bound on first category for quantile binning
- fix handling of numeric features names in pandas for error analysis tree view
- reverse y axis in heatmap so that higher values are towards the top of the matrix
- add warnings for 2d heatmap case and test for quantile binning with duplicate edges
-
Causal
- sort causal data point by effect size
- set treatment list spin range
- fix causal global effect plot x-axis titles cut off
- fix causal categorical treatment average gains text
-
Counterfactual
- change current class label for regression task
- raise exception when users configures less than required counterfactuals
- show options for categorical features on counterfactual list
-
Fairness
- add error handling for missing metrics in fairness
-
- other
- remove Axios library from dependencies
- rename wrapped-flask to rai_core_flask
- remove deprecated babelconfig option in workspace.json
release raiwidgets 0.11.0
v0.11.0
- new features
- table for list of available cohorts
- added more metrics to ErrorAnalysisDashboard, including precision, recall and mean abs error
- bug fixes
- fix various UI usability bugs
- fix for clicking custom scatter plot point
-
Error Analysis
- fix selecting what if datapoint crashing dashboard
- fix error in quantile binning when there are too many duplicate points
- fix displayed error rate in all data cohort in cohort list
- fix error in heatmap for classification case
- fix features list not getting disabled in static view
- update description text for static view
- other
- improve validation for model prediction
- reorder counterfactual table columns
- add absolute sorting for local importance charts
- improve spacing and fonts across dashboard
- fix typos in docstrings
- improve wording of text on dashboard
- add warnings for soft data limits
- remove font override for all components
- add code coverage to repository
release raiwidgets 0.10.0
v0.10.0
-
educational materials
- update current model analysis classification notebook by adding treatment features for causal analysis
-
new features
-
breaking changes
-
bug fixes
-
Model Analysis:
- change icon for ViewList in main menu
- fix model statistics component not showing all cohorts
- put causal component last, remove navigation
- hide removal dialog on click
- move cohort info and button to sticky menu, replace error stats with filters
-
Error Analysis:
- refactor heatmap code and fix display issues for categoricals
- fix incorrect values in heatmap when there are empty error cells
- added quantile binning and number of bins functionality to erroranalysis package
- added quantile binning and number of bins to UX and hooked up calls to backend
- added tooltip to list of features for tree view
- fix backend to support equals filter for integer categorical columns
- fix backend to support include and exclude filters for integer categorical columns
- update default value for matrix feature
- fix cohort when switching tree view
-
Causal:
- remove unnecessary z-index on what-if and causal components
- update causal policy to support categorical features
- format policy tree categorical features
- fix description on how confounding features are handled
- use lighter lines for causal plotly point graphs
-
Fairness:
- incorporate uncertainty quantification in fairness widget v2
- add dropdown to dataset explorer by default
- fix removing what if point will keep a faded start on the chart
- suppress distracting logging from flask when running widgets
- extend unseen categorical check to counterfactuals
- upgrade dice-ml to 0.7
- update counterfactual list to allow string values
- fix large x and y axis going out of bounds
- refactor multiple instances of convert to list method in several dashboards
- add missing interpretability information and guidelines
-
-
other
- fix message and check for environment in release pipeline
release raiwidgets 0.9.1
v0.9.1
- educational materials
- new features
- breaking changes
- error report in erroranalysis package had json_tree renamed to tree and json_matrix renamed to matrix
- bug fixes
- fix econML inputs
- add categorical feature to causal whatif
- fix counterfactual panel scroll bar
- round precision in causal analysis tables
- fix calculating feature importances after error report due to dataframe on analyzer being modified
- other
- dump information about the most time consuming tests
release raiwidgets 0.9.0
v0.9.0
- educational materials
- add Model Analysis notebook for regression
- new features
- local importance chart for counterfactual dashboard
- add multi causal policy support
- add delete tab dialog and remove inline widget for model assessment dashboard
- add treatment policy and table to causal analysis dashboard
- create new custom individual feature importance component
- breaking changes
- bug fixes
- explicitly pass model task to MimicExplainer in model analyzer
- validate treatment features in causal manager
- Model Analysis:
- sticky layer for cohort and dashboard settings
- restrict EA tree to 80% width
- restrict ipykernel<0.6
- fix flyout title and description
- fix add button behavior by adding components in the appropriate spot
- show add button only for components that can render
- Error Analysis:
- remove 100k row limit
- fix zero nodes in tree view
- fix search functionality in features list
- fix displayed filter order in tree view to start from root node
- raise warning for older version of pandas when feature names contain '-'
- expand number of decimals shown in filter tooltips when values below 0.1
- fix regression case for 1d heatmap matrix metrics being calculated incorrectly
- set original data from dropdown for counterfactuals
- use test data instead of train data for error analysis manager in model analyzer
- causal analysis to keep all levels in dataframe outputs
- pin networkx pip package to prevent installation of matplotlib
- add target_column causal policy gain to UI control
- other
- unit tests for multiclass classification for causal analysis
- add validations for input parameters for Model Analysis class
- add model analysis notebooks to notebook gate
- add test for causal manager for None categoricals
- fix example code for model assessment
release raiwidgets 0.8.0
v0.8.0
- educational materials
- new features
- enabled categoricals on MimicExplainer for the explainer manager
- updated causal constants to enable linear models and allow high cardinality features by default
- support newest policy output format from econml
- breaking changes
- continuous_features in CounterfactualManager.add() is deprecated in favor of categorical_features in ModelAnalysis
- bug fixes
- fixed error analysis add after model analysis deserialization
- other
- improved tests for visualization dependencies
release raiwidgets 0.7.0
v0.7.0
- educational materials
- new features
- added widget serialization for causal analysis
- breaking changes
- bug fixes
- other
release raiwidgets 0.5.0
v0.5.0
- educational materials
- cleanup commented out explainers from interpretability notebook
- new features
- added support for causal, counterfactual, and error analysis tools in responsibleai
- added model analysis widget to raiwidgets
- added support for regression to error analysis
- added policy tree to causal manager
- breaking changes
- renamed package raitools to responsibleai
- renamed top-level class RAIAnalyzer to ModelAnalysis
- removed support for fairness in responsibleai package
- bug fixes
- fixed bug in precision statistic calculation in dashboard
- fixed bug in loading explanations with ModelAnalysis
- fixed on predict method to use correct dtypes
- fixed individual importances chart when no data available
- other
- create combined feature importances tab for local and global explanation in model assessment
- add section headers to model assessment dashboard
- add manual cohort creation to the model assessment dashboard