LoginSign Up
← Back to all sparks
P

poissonreg

ANALYTICS
Velocity0.0

parsnip wrappers for Poisson regression. Contribute to tidymodels/poissonreg development by creating an account on GitHub.

github.com/tidymodels/poissonreg ↗

poissonreg gave its models away to parsnip and kept the glue — now it just keeps glmnet honest.

tidymodelscount-regressionglmnetr-languagemaintenance
◆Current state
poissonreg is a tidymodels extension that wires Poisson and zero-inflated count regression into the parsnip interface. Its defining event was giving up ownership: the model definition functions moved into parsnip itself, leaving this package as engine bindings and prediction plumbing. The current dev release is entirely correctness and hygiene work — glmnet predictions now default to mean counts rather than the linear predictor, and single-observation prediction works at last.
◆Where it's heading
Release cadence has collapsed from yearly to a four-year gap between 1.0.1 and the current development version, and the content has shifted from features to deduplication against parsnip — copied helper functions replaced by the upstream originals, obsolete generic registrations removed, tests migrated to the shared extension-package pattern. This is what a stabilized tidymodels satellite looks like: the interface lives upstream, and the package's job is to not drift from it.
◆Prediction
The dev version's accumulated fixes point to a CRAN release of 1.0.2 as the next move, with content limited to the glmnet prediction corrections rather than any new engine or model type.

◆Recent moves

  1. 3mo ago

    glmnet predictions now default to mean counts

    A correctness release for the glmnet engine: `multi_predict()` had been returning the linear predictor where mean counts were expected, and single-observation prediction simply did not work. The rest is deduplication against parsnip — copied helpers, obsolete generic registrations, and bespoke test patterns all replaced by the upstream versions.

    View source ↗
  2. 3y ago

    Documentation regenerated for valid HTML5

    A CRAN-hygiene release with no functional change — `.Rd` files regenerated so the rendered help pages validate as HTML5. The same box-ticking exercise that passed through most of CRAN in the same period.

    View source ↗
  3. 4y ago

    Case weight support tracks parsnip 1.0.0

    Brings the package in line with case weights as introduced in parsnip 1.0.0. A downstream adaptation rather than an independent feature — the pattern that defines this package's role after the model definitions moved upstream.

    View source ↗
  4. 4y ago

    Model definitions move out of poissonreg into parsnip

    ⚡ SPARK

    One line of changelog that redefines the package: `poisson_reg()` and its siblings now live in parsnip, leaving poissonreg as engine bindings around models it no longer owns. Everything since has been keeping those bindings synchronized with upstream.

    View source ↗
  5. 4y ago

    glm becomes the default engine; tidy() for hurdle models

    Gives `poisson_reg()` a working default engine instead of requiring an explicit choice, and adds `tidy()` methods for the pscl hurdle and zero-inflated fitters. Small surface, but it is the release that made zero-inflated counts legible to the broom-style tidying pipeline.

    View source ↗
  6. 5y ago

    First release, with a glmnet column-order safeguard

    The initial release, notable for working around a glmnet bug where a different column order silently produced wrong predictions. Silent-wrong-answer defenses on the glmnet engine have been a recurring thread in this package ever since.

    View source ↗