Research map / Learning theory and algorithms / Prediction and decision losses
Explaining the Success of AdaBoost and Random Forests as Interpolating Classifiers
Abraham J. Wyner, Matthew Olson, Justin Bleich, David Mease
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| boost deep trees (2^8 nodes) instead of stumps (ablation) | AdaBoost · 5D simulation, 250 iterations | points differing from Bayes rule (of 1000): 141/1000 -> 116/1000 | Improved |
| run AdaBoost for more iterations beyond interpolation (10 to 100) (ablation) | AdaBoost · 5D simulation, 200 repetitions | disagreement with Bayes rule: .19 -> .15 | Improved |
| run AdaBoost for 1000 instead of 9 iterations (ablation) | AdaBoost · 20D pure noise | proportion of hold-out classified +1 (error): .1433 -> .0008 | Improved |
| average ten interpolating sub-ensembles (100 to 1000 iterations) (ablation) | AdaBoost · 20D pure noise | agreement with Bayes rule: 80% (single interpolator) -> 99.92% | Improved |
| use AdaBoost or random forest instead of one-NN as interpolator | interpolating classifiers · 2D pure noise (p=0.8) | fraction classified +1: 79% -> 87% (AdaBoost), 94% (random forests) | Improved |
| use AdaBoost or random forest instead of pruned CART or one-NN | classifiers · 2D circle with signal | overall error rate: 0.18 (CART), 0.20 (one-NN) -> 0.13 | Improved |
| flip 100 training labels (about 2%) | AdaBoost depth-8 trees · Phoneme | test error: 9.0% -> 10.2% | Hurt |
| flip 100 training labels (about 2%) | Random forest · Phoneme | test error: 9.4% -> 10.0% | Hurt |
| flip 100 training labels (about 2%) | One-NN · Phoneme | test error: 10.5% -> 12.6% | Hurt |
| increase depth of boosted trees (ablation) | AdaBoost · Phoneme | test error | Improved |
| add 5% label noise | AdaBoost vs one-NN · Haberman | increase in test error: 1.55 (one-NN) -> 0.13 (AdaBoost) | Improved |
| add 5% label noise | AdaBoost vs one-NN · German credit | increase in test error: 0.68 (one-NN) vs 0.29 (AdaBoost), 0.68 (RF) | No clear effect |
Closest papers by meaning
- Towards Convergence Rate Analysis of Random Forests for Classification (NeurIPS 2020 · similarity 0.88)
- Randomization as Regularization: A Degrees of Freedom Explanation for Random Forest Success (JMLR 2020 · similarity 0.88)
- Precision-based Boosting (AAAI 2021 · similarity 0.88)
- Pushing the Envelope of Gradient Boosting Forests via Globally-Optimized Oblique Trees (CVPR 2022 · similarity 0.88)
- A Boosting-Type Convergence Result for AdaBoost.MH with Factorized Multi-Class Classifiers (NeurIPS 2024 · similarity 0.87)
- The Many Faces of Optimal Weak-to-Strong Learning (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
