Research map / Label-efficient and robust learning / Recognition and long-tail
Data-faithful Feature Attribution: Mitigating Unobservable Confounders via Instrumental Variables
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes IV-SHAP and IV-IG, which train a confounder-free model on instrumental-variable re-estimated features so that SHAP/IG attributions are faithful to the data-generation process under unobservable confounders.
Problem
Feature attribution (SHAP, IG) applied to predictive models trained on E[y|x] absorbs unobservable confounder effects and attributes them to correlated observed features, yielding attributions unfaithful to the data-generating process.
Approach
Train a model on the expected target under instrumental-variable re-estimated features (fitting E[y|x] with the confounded feature integrated over an IV-based conditional distribution), so the confounder term becomes a constant and SHAP/IG attributions reflect data-generation contributions.
Key results
- On Griliches76, IV-SHAP reduces relative error from about 0.55-0.57 (SHAP) to 0.146-0.184 across years, and IV-IG from about 0.55 (IG) to 0.135-0.178.
- On synthetic datasets, IV-SHAP and IV-IG show significantly smaller attribution errors than SHAP and IG across feature deviations and noise levels rho.
- Abstract reports up to 67% relative improvement over baselines in the error of attribution ratio on the real dataset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train the attribution model with instrumental variables (IV-SHAP instead of SHAP) | SHAP attribution · Griliches76, YEAR 1–5 | relative error: 0.566 ± 0.041 -> 0.184 ± 0.032; 0.569 ± 0.053 -> 0.162 ± 0.026; 0.569 ± 0.040 -> 0.172 ± 0.025; 0.548 ± 0.047 -> 0.157 ± | Helped |
| Train the attribution model with instrumental variables (IV-IG instead of IG) | Integrated Gradients attribution · Griliches76, YEAR 1–5 | relative error: 0.554 ± 0.044 -> 0.178 ± 0.025; 0.582 ± 0.052 -> 0.152 ± 0.020; 0.583 ± 0.044 -> 0.165 ± 0.028; 0.467 ± 0.047 -> 0.149 ± | Helped |
| Use instrumental-variable attribution rather than SHAP | SHAP attribution · Synthetic Datasets A and B | attribution error | Helped |
| Use instrumental-variable attribution rather than IG | Integrated Gradients attribution · Synthetic Datasets A and B | attribution error | Helped |
| Use instrumental variables to train the confounder-free attribution model | Feature attribution methods · Real dataset | error of attribution ratio: up to a 67% relative improvement | Helped |
Surprising findings
- On the Angrist-Krueger data, IV-SHAP and IV-IG yield attribution ratios close to the plain SHAP/IG values (0.0206 and 0.0207 vs 0.0218), a small difference.
Limitations
- Requires suitable instrumental variables for features affected by unobserved confounders, which may be hard to find.
- Theoretical derivations assume linear influence of unobserved confounders on target.
Experimental setup
Datasets: Synthetic datasets A and B (generated) Griliches76 Angrist and Krueger census dataset
Benchmarks: Synthetic datasets A and B Griliches76 (relative error of attribution ratio) Angrist and Krueger (attribution ratio comparison)
Metrics: absolute attribution error vs ground truth relative error of average attribution ratio
Baselines: SHAP IG
Builds on: SHAP Integrated Gradients two-stage instrumental variable estimation causal Shapley values
Closest papers by meaning
- Regional Explanations: Bridging Local and Global Variable Importance (NeurIPS 2025 · similarity 0.86)
- Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation (ICML 2024 · similarity 0.86)
- Intriguing Properties of Data Attribution on Diffusion Models (ICLR 2024 · similarity 0.86)
- Taming Hyperparameter Sensitivity in Data Attribution: Practical Selection Without Costly Retraining (NeurIPS 2025 · similarity 0.85)
- quanda: An Interpretability Toolkit for Training Data Attribution Evaluation (NeurIPS 2026 · similarity 0.85)
- Provably Better Explanations with Optimized Aggregation of Feature Attributions (ICML 2024 · similarity 0.85)
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.