Research map / Learning theory and algorithms / Prediction and decision losses
Being Properly Improper
Tyler Sypherd, Richard Nock, Lalitha Sankar
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 |
|---|---|---|---|
| Replace log-loss (alpha=1) boosting with alpha-loss (alpha in {1.1,2,4}) via PM-Boost | PM-Boost on xd6 with feature noise · xd6 feature noise p=0.5 | accuracy: 0.96 vs 0.88 AdaBoost, 0.83 XGBoost at p=0.5 | Improved |
| Larger alpha (1.1 to 4.0) in PM-Boost (ablation) | PM-Boost on xd6 feature noise · xd6 feature noise p=0.5 | accuracy: 0.91 at alpha=1.1 and 2.0 to 0.96 at alpha=4.0 | Improved |
| Alpha-loss boosting vs AdaBoost/XGBoost under no or low feature noise | PM-Boost on xd6 · xd6 feature noise p=0 to 0.25 | accuracy: all methods at or near 1.00 | No clear effect |
| PM-Boost with alpha-loss on noisy class labels (symmetric label noise) versus proper log-loss boosting | PM-Boost on cancer dataset with class noise · cancer class noise 0/15/30% | accuracy: qualitative gains over AdaBoost and XGBoost when twist present | Improved |
| Adaptive alpha (Menon PM-Boost, alpha estimated from data) versus fixed alpha=1.1 (ablation) | PM-Boost on xd6 with class noise · xd6 class noise | accuracy: slightly better in the paper's words | Mixed |
| Focal loss generalization of log-loss | focal loss · theory | twist-properness | Hurt |
| Super Loss wrapper | Super Loss · theory | twist-properness | Hurt |
Closest papers by meaning
- LegendreTron: Uprising Proper Multiclass Loss Learning (ICML 2023 · similarity 0.86)
- Lower-Bounded Proper Losses for Weakly Supervised Classification (ICML 2021 · similarity 0.85)
- Realizable $H$-Consistent and Bayes-Consistent Loss Functions for Learning to Defer (NeurIPS 2024 · similarity 0.84)
- Loss factorization, weakly supervised learning and label noise robustness (ICML 2016 · similarity 0.84)
- When Does Optimizing a Proper Loss Yield Calibration? (NeurIPS 2023 · similarity 0.84)
- Robust Loss Functions for Training Decision Trees with Noisy Labels (AAAI 2024 · similarity 0.83)
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.
