Research map / Label-efficient and robust learning / Recognition and long-tail
Binary Classification from Positive-Confidence Data
Takashi Ishida, Gang Niu, Masashi Sugiyama
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 |
|---|---|---|---|
| Use the Pconf risk objective instead of naive confidence-weighted classification | linear classifier · Synthetic Gaussian setups A-D | classification accuracy: 89.7\pm0.6 vs 88.7\pm1.2; 81.2\pm1.1 vs 78.1\pm1.8; 90.2\pm9.1 vs 82.7\pm13.1; 91.5\pm0.5 vs 90.8\pm0.7 | Improved |
| Use Pconf instead of naive confidence-weighted classification | three-hidden-layer neural network · Fashion-MNIST binary class pairs | classification accuracy: 92.14\pm4.06 vs 85.30\pm9.07; 96.00\pm0.29 vs 96.08\pm1.05; 91.52\pm1.14 vs 89.31\pm1.08; 98.12\pm0.33 vs 98.13\pm1.12; | Mixed |
| Use Pconf instead of naive confidence-weighted classification | CNN · CIFAR-10 airplane paired with other classes | classification accuracy: 82.68\pm1.89 vs 76.21\pm2.43; 82.23\pm1.21 vs 80.66\pm1.60; 85.18\pm1.35 vs 89.60\pm0.92; 87.68\pm1.36 vs 87.24\pm1.58; | Mixed |
| Add Gaussian noise to positive-confidence values, increasing noise standard deviation (ablation) | Pconf classification · Synthetic setups A-D | classification accuracy: Setup A: 89.8\pm0.6 -> 85.9\pm2.5; Setup B: 81.2\pm0.9 -> 77.8\pm1.4; Setup C: 92.4\pm1.7 -> 88.0\pm9.5; Setup D: 91.6\p | Hurt |
| Use naive confidence-weighted risk instead of the proposed Pconf risk | Pconf empirical risk minimization · Theoretical analysis | risk consistency | Hurt |
Closest papers by meaning
- Learning from Positive and Unlabeled Data with Arbitrary Positive Shift (NeurIPS 2020 · similarity 0.87)
- Learning from Similarity-Confidence Data (ICML 2021 · similarity 0.86)
- Classification from Positive, Unlabeled and Biased Negative Data (ICML 2019 · similarity 0.86)
- Surprises in Proper Positive-Only Learning (NeurIPS 2026 · similarity 0.85)
- Positive and Unlabeled Learning with Controlled Probability Boundary Fence (ICML 2024 · similarity 0.84)
- Learning from positive and unlabeled examples -Finite size sample bounds (NeurIPS 2025 · similarity 0.84)
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.
