LoginSign Up
← Back to all sparks
H

hstats

ANALYTICS
Velocity0.0

Friedmans H-statistics

github.com/ModelOriented/hstats ↗

hstats settled into maintenance after its 1.0 restructuring, with model coverage the only thing still growing.

interaction statisticspartial dependencemodel explainabilityr packagemodeloriented
◆Current state
hstats computes Friedman's H-statistics, partial dependence, ICE curves and permutation importance for any model exposing a prediction function. The releases in view are consolidation: performance work on plain data.frames, ICE facetting for multioutput models, ranger survival support, and a ggplot 4.0 compatibility pass in 2025. The package moved to the ModelOriented organisation in 1.2.0.
◆Where it's heading
The structural work — the hstats_matrix object, quantile approximation, revised plotting — landed in 1.0.0 just outside this window, and nothing since has changed the package's shape. What continues is model-coverage plumbing: mlr3 classification modes, ranger survival behind a survival argument, and factor predictions added in 1.1.0 then removed again in 1.2.0. The most recent releases are compatibility-driven, tracking ggplot2 rather than the interaction statistics.
◆Prediction
Expect the next release to be another dependency-compatibility pass or a new model backend working out of the box, rather than new interaction statistics.

◆Recent moves

  1. 10mo ago

    ggplot 4.0 compatibility and test coverage

    A compatibility release: a ggplot 4.0 fix, a reverse-dependency check, more tests and README timings. Nothing changes for callers beyond continuing to work.

    View source ↗
  2. 1y ago

    ranger survival models supported out of the box

    ranger survival models no longer need a hand-written prediction function, with a survival argument choosing between cumulative hazards and survival probabilities per time point. kernelshap gained the identical capability the same day.

    View source ↗
  3. 2y ago

    Moves to ModelOriented; factor predictions removed

    The repository moves under the ModelOriented organisation, and factor-valued predictions — added only eight months earlier — are dropped along with the classification_error loss. A narrowing of scope rather than an addition.

    View source ↗
  4. 2y ago

    ICE facets for multioutput models; mlr3 fixes

    Multioutput ICE plots switch from colour to facets, with swap_dim restoring the old behaviour. mlr3 non-probabilistic classification starts working, and probabilistic classification now requires an explicit predict_type.

    View source ↗
  5. 2y ago

    Faster data.frame paths; NaN H-statistics fixed

    Most functions get significantly faster on plain data.frames, and normalised H-statistics that could evaluate to 0/0 now return 0 instead of NaN in multivariate cases.

    View source ↗
  6. 2y ago

    Factor predictions and line-style 2D partial dependence

    Factor predictions become supported through one-hot encoding, and two-dimensional partial dependence gains a line view as an alternative to the heatmap. Both the one-hot path and general performance were sped up in the same release.

    View source ↗