LoginSign Up
← Back to all sparks
T

tidypredict

ANALYTICS
Velocity0.0

Translate fitted R models into SQL so predictions can run inside a database

github.com/tidymodels/tidypredict ↗

tidypredict now translates the gradient-boosting libraries people actually deploy

tidypredictsql-generationgradient-boostingin-database-scoringr-package
◆Current state
tidypredict converts fitted R models into SQL and dplyr expressions so predictions can run inside a database instead of in R. The 1.1.0 release added rpart, CatBoost, and LightGBM, with full objective and tree-type coverage for the boosted models. That followed 1.0.0, which broke random-forest output into a single formula, added glmnet, and cut fit-translation time for xgboost, partykit, and ranger.
◆Where it's heading
The package's value scales directly with how many model types it can translate, and the recent work has concentrated on the tree ensembles that dominate tabular modelling in practice. Coverage now extends past what parsnip wraps, since raw CatBoost models are supported alongside parsnip and bonsai ones with an explicit escape hatch for categorical features. Performance work on the translation step suggests the models being converted have grown large enough for that to matter.
◆Prediction
With the major boosting libraries covered, the remaining gap is what happens to preprocessing, so tighter integration with recipes or orbital for translating whole workflows is the natural next step.

◆Recent moves

  1. 5mo ago

    tidypredict 1.1.0 adds CatBoost, LightGBM, and rpart support

    ⚡ SPARK

    Support arrives for rpart decision trees, CatBoost, and LightGBM, with coverage of a dozen objectives and both oblivious and non-oblivious tree types. It closes the gap between what tidypredict can translate and what practitioners actually fit on tabular data.

    View source ↗
  2. 8mo ago

    tidypredict 1.0.1 fixes base_score extraction for xgboost 3

    Fixes base_score not being extracted correctly for xgboost version 3 and higher. A compatibility fix against an upstream change.

    View source ↗
  3. 8mo ago

    tidypredict 1.0.0 unifies random forest output and adds glmnet

    The 1.0.0 release makes ranger and randomForest emit a single formula instead of a list of expressions, adds glmnet support, extends xgboost to tweedie and count:poisson objectives, and speeds up fit translation across several engines. A breaking output change, but a narrow one aimed at making generated code easier to consume.

    View source ↗
  4. 1y ago

    tidypredict 0.5.1 exports internals for the orbital package

    Exports a number of internal functions for use by the orbital package. Plumbing for a downstream consumer.

    View source ↗
  5. 3y ago

    tidypredict 0.5 hands maintenance to a new maintainer

    A maintainer handover to Edgar Ruiz, with the stringr dependency removed and assorted XGBoost and earth fixes. Custodial rather than substantive.

    View source ↗
  6. 4y ago

    tidypredict 0.4.9 relicenses to MIT and fixes SQL generation

    Fixes GLM binomial earth models, adds single simple Cubist model handling, corrects SQL generation for binary:logistic xgboost objectives, and relicenses the package from GPL-3 to MIT. The relicensing matters more than the fixes for anyone embedding this in commercial tooling.

    View source ↗