Repository navigation
What is new
DistilledFIGS(teacher=None, max_rules=256, synthetic=0): a FIGS model (imodels) fitted to a teacher's class probabilities instead of the labels, withpredict_proba,predictandfidelity(X). Install withpip install neural-trees[figs].
Distilling into FIGS is not a new method (FIGS-BD, arXiv 2503.06730; imodels.DistilledRegressor). It is added because the equal-size audit behind the neural-trees paper measured it on held-out test rows, with the hypotheses fixed before the run (9 OpenML tables x 3 seeds, tuned XGBoost teacher):
| splits | FIGS on labels, acc minus teacher | AUC | DistilledFIGS, acc minus teacher | AUC |
|---|---|---|---|---|
| 128 | -2.86 | 86.44 | -2.28 | 87.41 |
| 256 | -3.06 | 85.79 | -1.93 | 87.80 |
| 512 | -3.58 | 84.31 | -1.84 | 87.95 |
At 512 splits it is ahead of FIGS on labels at its best size (128) by +1.02 accuracy and +1.52 AUC, on all 9 tables. A full EBM with interactions stays about one point ahead.
Packaging
- On Python 3.9 the
figsextra installs imodels below 3, which does not import there; it also installssix, which imodels 3.0.4 imports without declaring it. - CI installs the
figsextra, so the DistilledFIGS tests run on Python 3.9 to 3.13.