LoginSign Up
← Back to all sparks
T

treeshap

ANALYTICS
Velocity0.0

Compute SHAP values for your tree-based models using the TreeSHAP algorithm

github.com/ModelOriented/treeshap ↗

treeshap keeps widening its tree-model coverage while the SHAP math stays put.

shapmodel explainabilitytree ensemblesr packageadapter coverage
◆Current state
treeshap computes exact SHAP values for tree ensembles in R, reaching each modelling framework through a per-framework unify() adapter. Since returning to CRAN in 2023 it has added GPBoost, ranger survival forests and multi-output models to that adapter layer. Four releases in three years, each dominated by adapter work contributed by users of one specific framework.
◆Where it's heading
The direction is breadth of model support rather than new explanation methods: every release since the first CRAN submission adds or repairs a unify() backend. Maintenance is community-driven, with named contributors fixing the framework they personally use. Nothing in these entries points at work on the SHAP algorithms themselves.
◆Prediction
Expect the next release to add or repair another unify() adapter as a contributor brings their own framework, rather than to change how explanations are computed.

◆Recent moves

  1. 3mo ago

    GPBoost support lands; xgboost adapter repaired

    GPBoost joins the adapter roster and the xgboost binding is repaired — the breadth-over-depth pattern that defines every treeshap release. The R floor moves to 4.1.0 as a side effect.

    View source ↗
  2. 2y ago

    Fixes broken lightgbm.unify examples

    A documentation-only patch repairing lightgbm.unify examples, issued shortly after the package's return from CRAN archival. No change to how explanations are computed.

    View source ↗
  3. 2y ago

    Multi-output model explanations added

    Multi-output models get their own unified and explanation classes, extending the adapter layer to a model shape it could not previously represent. Survival-forest handling from the prior release is corrected at the same time.

    View source ↗
  4. 2y ago

    First CRAN release consolidates the unify() adapters

    ⚡ SPARK

    The first CRAN release collapses the per-framework adapters behind one generic unify() and brings random survival forests into scope — the point at which treeshap became an installable general-purpose SHAP backend rather than a GitHub project.

    View source ↗