Partially Interpretable Decision Trees Defer to Black-Box
Original title: Partially Interpretable Decision Trees
Three algorithms for partially interpretable decision trees match or beat standard CART on real datasets.
Partially interpretable models extend interpretable models with deferral to a black-box for accuracy. The paper proposes three algorithms: exact-search Branches extension, a CART conversion, and a generalized CART with deferral-aware splitting. Performance is similar across datasets; the deferral-aware version outperforms on a synthetic inhomogeneous dataset.
- Exact search optimizes the tree
- Conversion runs on standard CART
- Generalized CART uses deferral-aware split
- Matches black-box accuracy when homogeneous
From the original article
1 min read at arxiv.orgPartially interpretable models extend simple interpretable models with the option to defer complicated inputs to an accurate but uninterpretable black-box model. By deferring more inputs, the model gets more accurate; by deferring fewer, it gets more interpretable. Existing procedures to train such partially interpretable models build on linear models or decision rules.
By Ludovic Schwartz · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.