LoginSign Up
← Back to all sparks
E

embed

ANALYTICS
Velocity0.0

Recipe steps that create embeddings and encodings of predictors for R models

github.com/tidymodels/embed ↗

embed keeps adding encoding steps while shedding its deep-learning dependencies

feature-engineeringrecipestidymodelsumapencoding
◆Current state
embed supplies recipes steps that turn categorical predictors into numeric representations — likelihood encoding, UMAP projection, string-distance collapsing. The 1.1.x line made UMAP arguments tunable and moved keras and tensorflow out of hard dependencies; 1.2.0 added analytical likelihood encoding with partial pooling and retired step_feature_hash() in favor of textrecipes.
◆Where it's heading
Two quiet directions run through these releases. One is making the steps tunable rather than fixed, so they participate properly in tidymodels grids. The other is boundary maintenance: heavy dependencies pushed to Suggests, overlapping steps handed to the package that owns them. Recent releases are thin and fix-driven.
◆Prediction
Expect further consolidation with textrecipes over which package owns which encoding step, and continued upkeep against xgboost and uwot releases rather than new step families.

◆Recent moves

  1. 6mo ago

    step_umap() zero-component bug fixed

    A single fix for UMAP computing zero components. No change in capability.

    View source ↗
  2. 8mo ago

    Compatibility with all xgboost versions

    Dependency compatibility work, invisible in normal use.

    View source ↗
  3. 11mo ago

    step_lencode() adds analytical likelihood encoding with pooling

    The most substantive release in this window: a new encoding step with a smooth argument for partial pooling on numeric outcomes. step_feature_hash() is fully deprecated in favor of textrecipes, continuing the boundary cleanup.

    View source ↗
  4. 1y ago

    UMAP initial and target_weight become tunable

    Turns two UMAP arguments into tunable parameters so they can participate in a tidymodels grid, and moves all messaging to cli. Small but on the package's main axis.

    View source ↗
  5. 2y ago

    keras and tensorflow moved to Suggests

    Cuts the install weight for everyone not using the deep-learning steps, and adds the UMAP arguments that 1.1.5 later makes tunable. Documentation for tidy methods is also reworked.

    View source ↗
  6. 2y ago

    step_collapse_stringdist() returns factors

    Changes the output type of a collapsing step to factors and repairs a regression that had broken multi-column likelihood encoding.

    View source ↗