Research map / Label-efficient and robust learning / Recognition and long-tail
Feature attribution methods
Formalizes and tests feature attribution through necessity, sufficiency and gradient-based tests.
51 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Recognition and long-tail, keyword cluster “explanations · concept · interpretable”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
15 papers in 2023–24 and 10 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 1.2% to 0.8%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · Understanding Dominant Factors for Precipitation over the Great Lakes Region (AAAI 2016 · 1 citation)
- 2019 · Granger-Causal Attentive Mixtures of Experts: Learning Important Features with Neural Networks (AAAI 2019 · 35 citations)
- 2020 · Understanding Global Feature Contributions With Additive Importance Measures (NeurIPS 2020 · 202 citations)
- 2021 · The effectiveness of feature attribution methods and its correlation with automatic evaluation scores (NeurIPS 2021 · 24 citations)
- 2022 · Do Feature Attribution Methods Correctly Attribute Features? (AAAI 2022 · 94 citations)
- 2023 · CRAFT: Concept Recursive Activation FacTorization for Explainability (CVPR 2023 · 87 citations)
- 2024 · Less is More: Fewer Interpretable Region via Submodular Subset Selection (ICLR 2024 · 1 citation)
- 2025 · One Wave To Explain Them All: A Unifying Perspective On Feature Attribution (ICML 2025 · 0 citations)
- 2026 · ShapBPT: Image Feature Attributions Using Data-Aware Binary Partition Trees (AAAI 2026 · 2 citations)
Earliest papers here
- Understanding Dominant Factors for Precipitation over the Great Lakes Region (AAAI 2016 · 1 citation)
- Granger-Causal Attentive Mixtures of Experts: Learning Important Features with Neural Networks (AAAI 2019 · 35 citations)
- Disentangling Influence: Using disentangled representations to audit model predictions (NeurIPS 2019 · 17 citations)
What has been tried
41 papers here report 370 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 9%, hurt 21%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add confidence score | CUB-200-2011 validation · Deletion AUC | 0.0811 -> 0.0613 | Helped | Less is More: Fewer Interpretable Region via Submodular Subset Selection (ICLR 2024 · 1 citation) |
| remove the sufficiency module | Citeseer and Pubmed · FID+ | Hurt | Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation (ICML 2024 · 1 citation) | |
| Use a uniform baseline rather than a zero baseline for Integrated Gradients | ImageNet-1000 evaluation split · IDSDS | 0.196 -> 0.255 | Helped | Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · 1 citation) |
| Add missingness interaction terms rather than use indicators only | FICO and Breast Cancer with synthetic MAR missingness · accuracy (FICO); AUC (Breast Cancer) | Helped | Interpretable Generalized Additive Models for Datasets with Missing Values (NeurIPS 2024 · 1 citation) | |
| compute attributions from the mel spectrogram instead of wavelet coefficients | ESC-50 fold 1 · Insertion | 0.449 -> 0.169 | Hurt | One Wave To Explain Them All: A Unifying Perspective On Feature Attribution (ICML 2025 · 0 citations) |
| use optimal clusters based on model structure rather than default clustering | 1st-order synthetic · TP | 0.97 -> 1.00 | Helped | Regional Explanations: Bridging Local and Global Variable Importance (NeurIPS 2025 · 0 citations) |
| Use the fold-change kernel-fitting variant instead of the linear-classifier variant | Almost all evaluated settings · localization performance | Helped | Prospector Heads: Generalized Feature Attribution for Large Models & Data (ICML 2024 · 0 citations) | |
| Optimize using probabilities rather than logits | ImageNet-1k, zero reference · LeRF–MoRF probability AUC difference | 24.98 -> 31.69 | Helped | Benchmarking Deletion Metrics with the Principled Explanations (ICML 2024) |
What reviewers will expect
Datasets and baselines named in the 14 papers of this approach from 2024–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Fashion-MNIST · 3 of 14 papers
- CIFAR10 · 2 of 14 papers
- MNIST · 2 of 14 papers
- ImageNet · 2 of 14 papers
Baselines
- Grad-CAM · 4 of 14 papers
- LIME · 4 of 14 papers
- Saliency · 3 of 14 papers
- RISE · 3 of 14 papers
- SmoothGrad · 3 of 14 papers
- Integrated Gradients · 3 of 14 papers
- IG · 2 of 14 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 25 papers; from OpenAlex and the papers' affiliations).
- The University of Western Australia · 2 papers
Use it
Download BibTeX of 40 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Self-explaining and visual explanation networks (108 papers)
- Prototype and concept dictionary models (104 papers)
- Post-hoc explanation analysis (86 papers)
- Concept bottleneck models (30 papers)
Elsewhere in the topic
- Multilingual language representation pretraining (507 papers)
- Contrastive representation theory (495 papers)
- Dense features and foundation priors (482 papers)
- Correspondence and causal representation learning (447 papers)
- Data augmentation and generalization bias (409 papers)
- Partial and multi-label learning (345 papers)
- Loss design and classifier theory (273 papers)
- Group-robust classification (251 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).