Research map / Label-efficient and robust learning / Recognition and long-tail
Non-splitting Neyman-Pearson Classifiers
Jingming Wang, Lucy Xia, Zhigang Bao, Xin Tong
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 |
|---|---|---|---|
| Use all training data (non-splitting) for both scoring and threshold instead of sample splitting | eLDA vs pNP-LDA · Toy LDA example (n0=n1=50) | type II error: .7638 -> .4478 | Improved |
| Use non-splitting eLDA instead of pNP-LDA | lung cancer dataset · Lung cancer dataset (alpha=0.01, delta=0.05) | type II error: 1 -> .104 | Improved |
| Use non-splitting eLDA instead of pNP-LDA | multi-tumor cancer dataset · Cancer dataset (n0=83, n1=91) | type II error: 1 -> .437 | Improved |
| Use eLDA instead of splitting NP methods on observed violation rate | eLDA · Simulation Example 1a | type I error violation rate: pNP-LDA largest 0.007 vs eLDA around delta=0.1 | Improved |
| Use the fixed-dimension variant feLDA instead of eLDA as p grows to 20 and beyond | feLDA · Simulation Example 1c | type I error control | Hurt |
| Lower the violation target delta from 0.1 to 0.05 and 0.01 for feLDA (ablation) | feLDA · Simulation Example 1c, 1c', 1c* | type I error as p increases | No clear effect |
| Lower delta from 0.1 to 0.05 for NP umbrella algorithm classifiers (ablation) | NP-LDA, NP-sLDA, NP-penlog, NP-svm · Simulation Example 1c vs 1c' | observed violation rate: k* = 61 for both delta=0.1 and delta=0.05 | No clear effect |
| Lower delta to 0.01 for NP umbrella algorithm classifiers (ablation) | NP umbrella methods · Simulation Example 1c* | observed violation rate: nu(61)=0.042 > 0.01 so threshold changes | Improved |
| Use eLDA over splitting methods at large n0 under heavy-tailed non-Gaussian data | eLDA/feLDA · Non-Gaussian simulations | type II error | Mixed |
| Increase dimension p for pNP-LDA (ablation) | pNP-LDA · Simulation Example 1c | type II error | Hurt |
| Fix n1=500 and vary n0 (Example 1b) instead of balanced sizes (ablation) | pNP-LDA · Simulation Example 1a vs 1b | largest violation rate: 0.007 -> 0.028 | Mixed |
Closest papers by meaning
- Neyman-Pearson classification: parametrics and sample size requirement (JMLR 2020 · similarity 0.95)
- Neyman-Pearson Classification under High-Dimensional Settings (JMLR 2016 · similarity 0.93)
- Neyman-Pearson Classification under Both Null and Alternative Distributions Shift (ICLR 2026 · similarity 0.85)
- Balancing Positive and Negative Classification Error Rates in Positive-Unlabeled Learning (NeurIPS 2025 · similarity 0.83)
- Concentration and excess risk bounds for imbalanced classification with synthetic oversampling (NeurIPS 2025 · similarity 0.82)
- The VOROS: Lifting ROC Curves to 3D to Summarize Unbalanced Classifier Performance (AAAI 2025 · similarity 0.80)
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.
