Research map / Label-efficient and robust learning / Noisy labels
Classification with Noisy Labels by Importance Reweighting
Tongliang Liu, Dacheng Tao
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 |
|---|---|---|---|
| add importance reweighting (IW) to hinge loss | kernel hinge loss · UCI benchmarks average (kernel) | accuracy (%): 75.04 -> 76.48 | Improved |
| add importance reweighting (IW) to logistic loss | kernel logistic loss · UCI benchmarks average (kernel) | accuracy (%): 71.75 -> 73.23 | Improved |
| add importance reweighting (IW) to hinge loss | linear hinge loss · UCI benchmarks average (linear) | accuracy (%): 75.03 -> 76.81 | Improved |
| add importance reweighting (IW) to logistic loss | linear logistic loss · UCI benchmarks average (linear) | accuracy (%): 75.05 -> 76.43 | Improved |
| estimate noise rates jointly (eIW) instead of cross-validation (IW) (ablation) | linear hinge IW · UCI benchmarks average (linear) | accuracy (%): 76.81 -> 75.72 | Hurt |
| estimate noise rates jointly (eIW) instead of cross-validation (IW) (ablation) | kernel hinge IW · UCI benchmarks average (kernel) | accuracy (%): 76.48 -> 76.29 | Hurt |
| estimate noise rates jointly (eIW) instead of cross-validation (IW) (ablation) | kernel logistic IW · UCI benchmarks average (kernel) | accuracy (%): 73.23 -> 73.13 | No clear effect |
| compare multiple kernel learning StPMKL against single-kernel hinge | kernel hinge loss · UCI benchmarks average (kernel) | accuracy (%): 75.04 -> 68.19 | Hurt |
| label-dependent costs (LD) versus plain hinge on Image at 0.4/0.4 noise | linear hinge loss · Image (linear) | accuracy (%): 73.74 -> 81.84 | Improved |
| use the proposed noise-rate estimator versus cross-validation | KLIEP-based estimator · Synthetic data (2D), noise (0.2, 0.2) | estimated noise rates: (0.424, 0.352) -> (0.213, 0.230) | Improved |
| use the proposed noise-rate estimator versus Scott's method | KLIEP-based estimator · Synthetic data (2D), noise (0.2, 0.2) | estimated noise rates: (0.176, 0.203) -> (0.213, 0.230) | No clear effect |
| use the proposed noise-rate estimator versus cross-validation | KLIEP-based estimator · Diabetes (UCI), noise (0, 0.4) | estimated noise rates: (0.288, 0.206) -> (0.026, 0.402) | Improved |
Closest papers by meaning
- Robust Classification via Regression for Learning with Noisy Labels (ICLR 2024 · similarity 0.91)
- Cost-Sensitive Learning with Noisy Labels (JMLR 2018 · similarity 0.90)
- Noisy Label Learning With Provable Consistency for a Wider Family of Losses (TPAMI 2023 · similarity 0.88)
- Learning from Noisy Labels with No Change to the Training Process (ICML 2021 · similarity 0.88)
- Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.87)
- Don’t Label Twice: Quantity Beats Quality when Comparing Binary Classifiers on a Budget (ICML 2024 · similarity 0.86)
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.
