Research map / Label-efficient and robust learning / Recognition and long-tail
Do Feature Attribution Methods Correctly Attribute Features?
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 SHAP rather than the other evaluated saliency maps | ResNet-34 bird classifiers · Bird species images with injected artifacts | Attr% in joint effective region: 69% at 40% | Mixed |
| Increase model test accuracy during training (ablation) | Saliency maps · Bird species images with injected artifacts | Attr% in joint effective region | Mixed |
| Increase manipulation visibility (ablation) | Saliency maps · Bird species images with injected artifacts | Attr% in joint effective region | Mixed |
| Increase original-label retention r (ablation) | Saliency maps · Bird species images with injected artifacts | Attr% on manipulated features | Mixed |
| Evaluate attention weights as feature attributions | Bi-LSTM with attention · BeerAdvocate, Article dataset | Attr% on article words: 8.6% | Hurt |
| Use the CN rather than the NC manipulation arrangement (ablation) | Bi-LSTM with attention · BeerAdvocate, CN and NC datasets | Attr% on non-correlating articles: 5.9% ± 4.0% vs. 12.6% ± 9.3% | Mixed |
| Train rationale models with a target selection rate instead of length regularization (ablation) | RL and CR rationale models · BeerAdvocate, Article dataset | Rationale precision and recall | Mixed |
| Set the rationale selection rate to its lowest tested value for the CR model (ablation) | CR rationale model · BeerAdvocate, Article dataset | Rationale precision and recall | Hurt |
| Select rationales with RL rather than CR training (ablation) | Rationale models · BeerAdvocate, CN and NC datasets | Precision on correlating articles | Hurt |
Closest papers by meaning
- Provably Better Explanations with Optimized Aggregation of Feature Attributions (ICML 2024 · similarity 0.87)
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.85)
- Data-faithful Feature Attribution: Mitigating Unobservable Confounders via Instrumental Variables (NeurIPS 2024 · similarity 0.85)
- Rethinking Robustness of Model Attributions (AAAI 2024 · similarity 0.85)
- Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation (ICML 2024 · similarity 0.85)
- quanda: An Interpretability Toolkit for Training Data Attribution Evaluation (NeurIPS 2026 · 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.