> ML_LITERATURE // NORI-2019-INTERPRETML-UNIFIED-FRAMEWORK-MACHINE-LEARNING-INTERPRETABILITY_v1.0
InterpretML: A Unified Framework for Machine Learning Interpretability
Harsha Nori, Samuel Jenkins, Paul Koch, Rich Caruana · arXiv preprint (2019)
systems2019industry-standardthirdPartyReproduced
Principal Contribution
Introduced Explainable Boosting Machines (EBM), combining modern tree boosting with GAMs and pairwise interaction detection.
Operational Relevance
Production-ready glassbox tool achieving full accuracy parity with XGBoost while providing exact intelligible attribution plots.
Assumptions
- Pairwise feature interactions capture the overwhelming majority of non-linear interaction energy in tabular data
Limitations
- Training scales quadratically with number of pairwise interaction candidates; cannot model 3-way interactions