Research map / Learning theory and algorithms / Prediction and decision losses
InstaSHAP: Interpretable Additive Models Explain Shapley Values Instantly
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.
Links SHAP and GAM through a functional ANOVA/variational view, proves SHAP-vs-GAM representation limits, and proposes InstaSHAP, a masked GAM training method that yields Shapley values in one forward pass.
Problem
SHAP explanations are slow to compute for black-box models and cannot represent feature interactions, especially under correlated inputs common in CV and NLP; prior variational SHAP formulations assumed independent features.
Approach
Unifies FastSHAP, FaithSHAP and GAM training under one variational least-squares framework over correlated inputs, and trains GAMs with Shapley-kernel masking so each additive component is included only when all its features are unmasked, yielding purified components whose sum gives Shapley values instantly.
Key results
- On tree cover, a low-dimensional GAM reaches 82.2% validation accuracy versus 80.4% for an MLP and 72.4% for GAM-1.
- On CUB, ResNet-50 reaches 65.0% fine and 81.8% coarse accuracy, while GAM-1 reaches only 33.2% fine and 53.7% coarse; GAM-3x3 reaches 46.8% fine and 66.8% coarse.
- On synthetic data, InstaSHAP shows lower MSE in approximating Shapley values than FastSHAP across training epochs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a low-dimensional GAM instead of GAM-1 | GAM on Bike Sharing · Bike Sharing | R²: 17.4% -> 6.23% | Helped |
| Use a low-dimensional GAM instead of GAM-1 | GAM on Tree Cover · Tree Cover | accuracy: 72.4% -> 82.2% | Helped |
| Use GAM-1 instead of ResNet-50 | CUB classifier · CUB fine-grained | accuracy: 65.0% -> 33.2% | Hurt |
| Use GAM-1 instead of ResNet-50 | CUB classifier · CUB coarse-grained | accuracy: 81.8% -> 53.7% | Hurt |
| Increase patch set from 1×1 to 2×2 (ablation) | GAM on CUB · CUB fine-grained | accuracy: 33.2% -> 45.8% | Helped |
| Increase patch set from 1×1 to 2×2 (ablation) | GAM on CUB · CUB coarse-grained | accuracy: 53.7% -> 66.3% | Helped |
| Increase patch set from 1×1 to 3×3 (ablation) | GAM on CUB · CUB fine-grained | accuracy: 33.2% -> 46.8% | Helped |
| Increase patch set from 1×1 to 3×3 (ablation) | GAM on CUB · CUB coarse-grained | accuracy: 53.7% -> 66.8% | Helped |
Surprising findings
- Standard 1D SHAP can miss a large portion of accuracy on CUB, with GAM-1 losing about 20 percentage points against ResNet-50 on fine-grained classification.
Limitations
- Low-order GAMs (GAM-1, GAM-4, GAM-9 on CUB) do not match the accuracy of deep CNNs in computer vision and NLP.
- The Stone-Hooker purification and higher-order extensions are only partially addressed; scaling beyond low-order interactions is not solved.
Experimental setup
Datasets: bike sharing tree cover CUB-200-2011 synthetic Gaussian data
Benchmarks: bike sharing tree cover CUB-200-2011 synthetic
Metrics: R^2 validation accuracy fine-grained accuracy coarse-grained accuracy MSE of SHAP approximation
Baselines: FastSHAP KernelSHAP MLP ResNet-50 GAM-1 GAM-2
Builds on: FastSHAP KernelSHAP FaithSHAP Sobol-Hoeffding decomposition GAM2
Closest papers by meaning
- Additive Models Explained: A Computational Complexity Approach (NeurIPS 2025 · similarity 0.88)
- SHAP values via sparse Fourier representation (NeurIPS 2025 · similarity 0.87)
- Helpful or Harmful Data? Fine-tuning-free Shapley Attribution for Explaining Language Model Predictions (ICML 2024 · similarity 0.87)
- SHAP Meets Tensor Networks: Provably Tractable Explanations with Parallelism (NeurIPS 2025 · similarity 0.86)
- shapiq: Shapley Interactions for Machine Learning (NeurIPS 2024 · similarity 0.85)
- Beyond TreeSHAP: Efficient Computation of Any-Order Shapley Interactions for Tree Ensembles (AAAI 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.