Unbiased MDI-like feature importance measure for random forests#31271

Closed

GaetandeCast wants to merge 30 commits intoscikit-learn:mainfrom

GaetandeCast:main

Commits on Apr 14, 2025

First working implementation of UFI, does not support multi output, has "extreme value" issues
Gaetan
committed

Commits on Apr 15, 2025

Removed the normalization inherited from the old MDI to avoid instability
Gaetan
committed
added multi output support
Gaetan
committed
removed redundant cross_impurity computations
Gaetan
committed

Commits on Apr 16, 2025

added mdi_oob
Gaetan
committed

Commits on Apr 17, 2025

redesigned ufi for better memory management
Gaetan
committed
removed a debug import
Gaetan
committed

Commits on Apr 18, 2025

added mdi_oob, cleaned the code
Gaetan
committed
better unified the code between ufi and mdi_oob
Gaetan
committed
fixed a call oversight
Gaetan
committed
fixed an error in mdi_oob computations
Gaetan
committed

Commits on Apr 22, 2025

changed tests on feature_importances_ to use unbiased FI too
Gaetan
committed

Commits on Apr 23, 2025

add tests to check that the added methods coincide with the papers and that they coincide with feature_importances_ on inbag samples
Gaetan
committed

Commits on Apr 24, 2025

Commits on Apr 25, 2025

Commits on Apr 28, 2025

Commits on Apr 29, 2025

adapted the tests to the previous commit
Gaetan
committed
Added log_loss ufi
Gaetan
committed
fixed the oob_score_ issue, simplified the self.value accesses
Gaetan
committed

Commits on Apr 30, 2025