New
FastRiskScoreClassifier: sparse integer risk scores. A few conditions, each worth a small number of points, with a calibrated risk for every total. It fits about 225× faster than FasterRisk by the median, with a lower training loss. See the post.- New blog index for the "Our favorite methods" posts.
Compatibility
- Tested on Python 3.10, 3.13 and 3.14, with the newest numpy, pandas, scikit-learn and numba.
MarginalShrinkageLinearModelRegressor is renamed MarginalShrinkageLinearRegressor; the old name still works, with a FutureWarning.- Importing
SLIMClassifier warns that it is deprecated in favour of FastRiskScoreClassifier.
- The numba-based models print a one-time notice before their first compile on a machine.
Fixes (from a pre-release bug sweep, each with a regression test)
- FastSmallTree:
time_limit / memory_limit are honoured- predictions follow the certified tree in float64
- exact tie-breaks under
balance / costs
- GPGam: correct predictions for features with a few uneven levels or a heavily tied minimum, and explicit settings are no longer overridden.
- All models:
- a refit on a numpy array removes a stale
feature_names_in_
- classifiers reject continuous targets
- clear errors for single-class targets
- Older models:
TaoTreeRegressor fits againGreedyRuleListClassifier(class_weight=...) takes effect- numpy-2 fixes in RF+ and BRL
DecisionTreeCCP* construct without arguments
- Packaging: the license now ships. The package no longer contains docs, notebooks or tests.
Full changes: #306, #307, #308
