Backlog

[interpret-ml]

This issue is to maintain links to all feature requests and bugs in one place. Issues listed here should have either the "enhancement " or "bug" label applied to them.

Issues not listed here are searchable with: is:issue is:open -label:bug -label:enhancement

Enhancements, python

• add method to convert EBMs to GradientBoostingRegressor (Getting Feature Importances of Individual Features from Interaction Terms #639)
• change smoothing round implementation so that it isn't impacted by feature order when used on nominals (Order of nominal feature levels change the results #629)
• per-feature smoothing parameters (Smoothing only for some features defined by the user #626)
• probability calibration plots (Probability calibration #625)
• GPU EBMs (Is GPU acceleration still in the works? #610)
• improve parallel handing of EBM data (Question: Parallel boosting? #519)
• add fully customizable python-based objectives (Does the EBM model support customized loss functions (objectives)? #498)
• confine pairwise interactions to the boundary of the convex hull (Extrapolation: confine 2-dimensional plots to convex hull of inputs? #458)
• add "warm_start" parameter to enable stage customization during fitting (Staged prediction/incremental fitting support #304). Also see issue # 403
• avoid resetting the index on call to show after explain_local (Is there a way to show explain_local without resetting the index? #72)
• user defined labels for local explanation samples (How to add index names--(strings such as ) #300)
• change boosting algorithm to stop it from producing graphs with large positive or negative scores in some regions (Can I limit magnitude of feature curves #122)
• ranking (Explaining learning to rank tasks #181)
• Dask arrays for large datasets (Large datasets: support for Dask arrays? #249)
• Spyder support for show function (Show(hist), Spyder3.8, testing on private data  #191)
• UI to show missing value scores (Doing some treatment for missing values #18)
• show error bars for interactions in the UI, perhaps with a hover (p-value of the functions of features/feature interactions #127)
• Improve contrast of interaction heatmaps (Improve contrast of interaction heatmaps #236)
• dashboard data disappears when changing tabs (Data shown in the Dashboard disappears when navigating across the tabs #106)
• add permutation Importance (Feature importance using Permutation #26)
• alternative feature importance techniques (Observing Alternative Feature Importance Techniques #374)
• add categorical support to Morris Sensitivity Analysis (PCA and Categoricals Interpretability Issue #275)
• alternate base estimators such as linear models, polynomial regression, or splines (Non-Treebased base models #185)
• Add waterfall plot (Add waterfall plot for local explanations #453)

Enhancements, R

• local explanations in R ([r] ebm_show local version? #222)
• R multiple new features (R: interpret package features #294)
• add threading to R and in general move the parallelism into C++ (Can't use n_jobs argument in R package #180)

Bugs

• predictions in R different from python (R: intercept may not be handled correctly #417)
• DataException: MissingColumnsInData, Expected column(s) 0 not found in fitted data (DataException: MissingColumnsInData, Expected column(s) 0 not found in fitted data #209)
• multiclass LIME (multiclass problem #306)
• LimeTabular seems to require Pandas dataframe for categoricals (Getting error while calling LimeTabular with ColumnTransformation  #322)
• Categorical features in PartialDependence (Categorical features in PartialDependence()? #217)
• show does not display any output in some settings (Show Displays Resulting in Blank Cell Output - Dash mode = "external"? #229)
• raise an exception when the multi-model dashboard is used inside the InlineProvider (Dashboard not showing results #333)
• in some scenarios showing multiple EBM dashboards could fail (Unable to show multiple Explanation Dashboard in the same notebook #350)