Understanding, Simplicity, And Decision Trees
summary
Shows that CART-style trees suffer a provable gap on smooth additive models by proving a lower bound using information theory that decision trees need many splits to mimic additive components. Builds a practical fix called FIGS: fast interpretable greedy tree sums that grows multiple trees with backfitting and a global split budget. Applied to clinical CT triage and public benchmarks, FIGS cuts model complexity (eight nodes down to five), raises AUC to about 0.81 on diabetes data, and yields roughly a 10% lift over single trees while remaining interpretable.
AI-written summary · about
chapters (7)
- 0:00Start
- 0:45Always Look for the Simplest Solution
- 4:04A Tribute to David Blackwell
- 5:37Proof of Word Theorem
- 22:32Bias Variance Decomposition
- 24:21Card Algorithm
- 33:10David Blackwell
related talks
40mCART: Not only Classification and Regression Trees
26mStop Explaining Black Box Models, Use Interpretable Models Instead
06mPërdorimi i Decision Trees dhe të dhënat për të marrë vendime
27mData-driven challenges in AI: scale, information selection, and safety
19mPrimo: Practical Learning-Augmented Systems with Interpretable Models
01h 28mHands-On with Tabular Foundation Models
29mSome Staged Tree Models For Learning From Interventions
15mInterpretable Machine Learning with R