Research map / Learning theory and algorithms / Prediction and decision losses
Data Shapley: Equitable Valuation of Data for Machine Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Remove negative-value images and reweight remaining images by positive Data Shapley value | Inception-V3 skin-lesion classifier · HAM10000 held-out set (400 images) | accuracy: 29.6% to 37.8% | Helped |
| Remove negative-value source images and emphasize positive-value images during training | Inception-Resnet-V1 gender classifier · PPB held-out set (400 images) | accuracy: 84.1% to 91.5% | Helped |
| Remove negative-value training examples and weight remaining examples by value | LSTM clinical-note classifier · Private Practice clinical notes | average prediction accuracy: 87.5% to 90.1% | Helped |
| Select new patients resembling high-value training patients instead of random patients | UK Biobank disease prediction · Breast and skin cancer prediction | prediction performance | Helped |
| Add patients resembling low-value training points | UK Biobank disease prediction · Breast and skin cancer prediction | prediction performance | Hurt |
| Remove low-value training points | UK Biobank logistic-regression disease predictor · Breast and skin cancer prediction | accuracy | Helped |
| Use Data Shapley rather than leave-one-out to rank mislabeled examples | Label-noise detection · Spam, flower classification, and Fashion MNIST | mislabeled-example detection | Helped |
| Increase image noise | Dog vs. Fish training images · Dog vs. Fish | average data value | Hurt |
| Use TMC-Shapley instead of leave-one-out to rank examples by their effect on model performance | Synthetic prediction tasks · Linear and third-order polynomial data | identification of points with positive performance effect | Helped |
| Use Gradient Shapley instead of TMC-Shapley (ablation) | Data valuation · Label-flip detection and patient data valuation | qualitative valuation results | No clear effect |
| Truncate sampled permutations after 25% of elements (ablation) | TMC-Shapley · Synthetic datasets | rank correlation with valuation without truncation: around 0.8 | Hurt |
| Use TMC-Shapley to estimate exact Shapley values | Logistic-regression valuation on synthetic data · Synthetic datasets (training sizes 4 to 14) | Pearson correlation with true Shapley value: 98.4% to 99.5% | Helped |
Closest papers by meaning
- A Distributional Framework For Data Valuation (ICML 2020 · similarity 0.92)
- DU-Shapley: A Shapley Value Proxy for Efficient Dataset Valuation (NeurIPS 2024 · similarity 0.91)
- Rethinking Data Shapley for Data Selection Tasks: Misleads and Merits (ICML 2024 · similarity 0.91)
- A Unified Framework for Provably Efficient Algorithms to Estimate Shapley Values (NeurIPS 2025 · similarity 0.88)
- Faithful Group Shapley Value (NeurIPS 2025 · similarity 0.87)
- Data Shapley in One Training Run (ICLR 2025 · 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.