LoginSign Up
← Back to all sparks
S

shapviz

ANALYTICS
Velocity0.0

SHAP Plots in R

github.com/ModelOriented/shapviz ↗

shapviz refines its SHAP plots release by release while chasing ggplot2's moving target.

shapvisualizationmodel explainabilityggplot2r package
◆Current state
shapviz turns SHAP values from XGBoost, LightGBM, H2O, kernelshap and other sources into standard diagnostic plots — importance, dependence, waterfall, force and interaction. Recent work is plot ergonomics: shared y-axis control across dependence plots, a bar view for interaction values, and axis collection via patchwork. The two most recent releases are pure compatibility and bug fixes.
◆Where it's heading
Two threads run in parallel here. One is visual refinement converging on conventions from Python's shap — the 0.10.0 notes openly float switching share_y to TRUE to match it. The other is connector maintenance, keeping pace with H2O, XGBoost 1.x and 2.x, shapr and permshap as each changes. Neither thread adds new explanation methods; shapviz's job is presentation, and it is being polished rather than extended.
◆Prediction
Expect share_y = TRUE to become the default and further ggplot2 4.x fallout, with connector updates arriving as the upstream SHAP packages release.

◆Recent moves

  1. 10mo ago

    ggplot 4.0 compatibility fix

    A two-pull-request release clearing a ggplot 4.0 compatibility warning ahead of CRAN submission. Sibling package hstats shipped the same fix the same day.

    View source ↗
  2. 1y ago

    Fixes duplicated bars in sv_interaction()

    A single bug fix: feature-pair bars appeared twice in the bar view added one release earlier. Small correction to a new feature rather than new capability.

    View source ↗
  3. 1y ago

    ggplot2 and patchwork dependency bumps

    Maintenance only — minimum versions raised for ggplot2 and patchwork, the two packages the plotting layer is built on.

    View source ↗
  4. 1y ago

    Shared y-axis control and bar-style interaction plots

    Dependence plots gain ylim and share_y so importance can be judged across several panels, and interaction values get a mean-absolute bar view. The notes flag a likely future default change to match Python's shap.

    View source ↗
  5. 1y ago

    H2O random forests gain TreeSHAP support

    H2O random forests become explainable via TreeSHAP for regression and binary classification, and background data for model-agnostic SHAP is easier to reach. Connector work driven by upstream H2O rather than by shapviz itself.

    View source ↗
  6. 1y ago

    Fixes a broken vignette link

    A documentation-only release correcting one link. No functional change.

    View source ↗