fastshap review
A focused R toolkit for feature attribution, with model-agnostic and tree-model workflows.
Reviewed by iTechGuides Editors · Editorial team · Updated Oct 2026
fastshap is an open-source R package for attributing supervised-model predictions to input features. It is aimed at R users who already work with fitted machine-learning models and need Shapley-based explanations. Rather than a hosted service, it runs in the user's environment, with Linux, macOS, and Windows listed as supported platforms. Its model-agnostic workflow relies on a prediction function supplied by the user, making the package applicable to supervised models that can be exposed through that interface.
The package offers approximate Shapley values calculated with Monte Carlo methods, plus exact Shapley values for linear models and Tree SHAP support for xgboost and lightgbm. Documentation also describes deterministic decomposition into main, total, and interaction effects. Optional parallel computation across features can help with computation, while raw per-simulation Shapley values provide a way to assess Monte Carlo uncertainty. Outputs are compatible with shapviz plotting functions, giving R users a visualization path without turning fastshap into a broader interpretability suite.
This is a narrow tool for feature attribution, not an end-to-end explanation or governance product: its listed capabilities do not include counterfactuals, fairness analysis, or audit reports. It is free and open source, with documentation available; distribution is through r-universe, and the documentation says it is no longer available on CRAN. Choose fastshap if Shapley explanations, especially for linear or supported tree models, are the work at hand and a user-managed R workflow fits. Consider another option if you need a hosted service or a wider explanation workflow.
fastshap pros and cons
- Where it wins
- Approximate Shapley values work with user-supplied prediction functions
- Exact values for linear models and Tree SHAP for xgboost and lightgbm
- Interaction decomposition and per-simulation values support deeper analysis
- Where it doesn't
- Focused on feature attribution, not a broader explanation workflow
- Requires fitted models and an R prediction function
- Distributed through r-universe; documentation says it is no longer on CRAN
fastshap fact sheet, pricing and score →
Advertiser disclosure: iTechGuides is reader-supported. Vendors can pay for top positions in our rankings and for a place on other products' pages, and we may earn a commission when you click some links. How we rank.
Last updated · How we research and update