Research map / Learning theory and algorithms / Prediction and decision losses
On the Accuracy of Influence Functions for Measuring Group Effects
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 |
|---|---|---|---|
| Use influence functions to estimate the effects of removing coherent groups instead of retraining | L2-regularized logistic regression · Constructed groups across Diabetes, Enron, Dogfish, CDR, and MNIST | Spearman correlation between predicted and actual effects: 0.89 to 0.99 | Helped |
| Use influence functions to estimate effects of natural groups | Logistic regression · MultiNLI crowdworker groups | Spearman correlation between predicted and actual effects on test loss: 0.77 to 0.86 | Helped |
| Remove groups with positive influence on test loss | Influence-function group-effect estimates · Constructed groups across five datasets | Agreement in direction and magnitude of test-loss effect | Mixed |
| Increase regularization strength | Influence-function group-effect estimates · Constructed groups across five datasets | Correlation between predicted and actual effects | Helped |
| Use labeling-function coverage as a signal of group influence | CDR labeling functions · CDR | Association with influence on test performance | Helped |
| Use labeling-function precision as a signal of group influence | CDR labeling functions · CDR | Association with influence on test performance | No clear effect |
| Increase the number of examples contributed by a crowdworker | MultiNLI crowdworker groups · MultiNLI | Influence on test performance | No clear effect |
| Remove the most prolific crowdworker's examples | MultiNLI logistic regression · MultiNLI | Overall test performance: 35,000 | Hurt |
| Approximate group effects with influence functions rather than retraining | Influence functions · Constructed groups across five datasets | Effect magnitude | Mixed |
Closest papers by meaning
- Rescaled Influence Functions: Accurate Data Attribution in High Dimension (NeurIPS 2025 · similarity 0.87)
- A Versatile Influence Function for Data Attribution with Non-Decomposable Loss (ICML 2025 · similarity 0.85)
- Influence Functions for Scalable Data Attribution in Diffusion Models (ICLR 2025 · similarity 0.84)
- Towards Robust Influence Functions with Flat Validation Minima (ICML 2025 · similarity 0.84)
- The Mirrored Influence Hypothesis: Efficient Data Influence Estimation by Harnessing Forward Passes (CVPR 2024 · similarity 0.84)
- Capturing the Temporal Dependence of Training Data Influence (ICLR 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.