Research map / Learning theory and algorithms / Prediction and decision losses
Prediction via Shapley Value Regression
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 ViaSHAP, which trains a model to output Shapley values per feature and obtains predictions by summing them, implemented with KANs and MLPs and evaluated on 25 tabular datasets and CIFAR-10.
Problem
Shapley-value explanations are computed post-hoc, adding inference-time cost; the paper asks whether a model can produce predictions directly from Shapley values while also yielding accurate explanations.
Approach
ViaSHAP makes the prediction the sum of learned per-feature Shapley values, trained jointly with a weighted least-squares Shapley loss over sampled coalitions plus the task loss, so explanations are produced during inference without a separate explainer.
Key results
- KAN-ViaSHAP obtains the highest average AUC rank over 25 tabular datasets, with differences from XGBoost and Random Forests statistically insignificant under Nemenyi test.
- KAN-ViaSHAP Shapley values give significantly higher similarity to KernelSHAP ground truth than FastSHAP on tabular data and images, according to the paper.
- Removing the link function is reported to significantly improve Shapley accuracy while keeping predictive performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KAN rather than an MLP approximator | ViaSHAP · 25 tabular datasets | AUC | Helped |
| Use ViaSHAP KAN rather than XGBoost or Random Forests | Tabular prediction · 25 tabular datasets | AUC | No clear effect |
| Train the KAN with Shapley loss (ablation) | KAN classifier · Tabular prediction | AUC | Helped |
| Increase beta (ablation) | ViaSHAP · Tabular datasets | Shapley value accuracy | Helped |
| Increase beta (ablation) | ViaSHAP · Tabular datasets | Predictive performance | No clear effect |
| Increase the number of sampled coalitions (ablation) | ViaSHAP · Tabular datasets | Predictive performance and Shapley value accuracy | No clear effect |
| Remove the link function (ablation) | ViaSHAP · Tabular datasets | Shapley value accuracy | Helped |
| Remove the link function (ablation) | ViaSHAP · Tabular datasets | Predictive performance | No clear effect |
| Remove the efficiency constraint (ablation) | ViaSHAP · Tabular datasets | Predictive performance and explanation accuracy | No clear effect |
| Use ViaSHAP KAN rather than FastSHAP for explanations | Shapley value approximation · Tabular data and images | Shapley value accuracy | Helped |
Surprising findings
- Removing the link function improves Shapley accuracy without hurting predictive performance.
- MLP-based ViaSHAP variants underperform all other competitors, even when parameter-matched to the KAN variant.
Limitations
- Hyperparameters chosen quasi-randomly and default settings used without tuning.
- Ground-truth Shapley values approximated by unbiased KernelSHAP run to convergence.
- MLP variants underperform; the approach is evaluated mainly on small tabular data and CIFAR-10 with models trained from scratch.
Experimental setup
Datasets: 25 public tabular datasets CIFAR-10
Benchmarks: 25 public tabular datasets (AUC) CIFAR-10
Metrics: AUC cosine similarity Spearman rank correlation R^2
Baselines: XGBoost Random Forests TabNet FastSHAP MLP-based ViaSHAP
Builds on: KernelSHAP FastSHAP Kolmogorov-Arnold Networks
Closest papers by meaning
- A Unified Framework for Provably Efficient Algorithms to Estimate Shapley Values (NeurIPS 2025 · similarity 0.87)
- Provably Accurate Shapley Value Estimation via Leverage Score Sampling (ICLR 2025 · similarity 0.87)
- SHAP values via sparse Fourier representation (NeurIPS 2025 · similarity 0.86)
- shapiq: Shapley Interactions for Machine Learning (NeurIPS 2024 · similarity 0.86)
- Regression-adjusted Monte Carlo Estimators for Shapley Values and Probabilistic Values (NeurIPS 2025 · similarity 0.86)
- OperatorSHAP: Fast and Accurate Shapley Value Estimation for Neural Operators (NeurIPS 2026 · 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.