Repository files navigation

embed has extra steps for therecipes package for embeddingpredictors into one or more numeric columns. Almost all of thepreprocessing methods aresupervised.

These steps are available here in a separate package because the stepdependencies,rstanarm,lme4, andkeras3, are fairly heavy.

Some steps handle categorical predictors:

• step_lencode_glm(),step_lencode_bayes(), andstep_lencode_mixed() estimate the effect of each of the factorlevels on the outcome and these estimates are used as the newencoding. The estimates are estimated by a generalized linear model.This step can be executed without pooling (viaglm) or with partialpooling (stan_glm orlmer). Currently implemented for numeric andtwo-class outcomes.

• step_embed() useskeras3::layer_embedding to translate theoriginalC factor levels into a set ofD new variables (<C).The model fitting routine optimizes which factor levels are mapped toeach of the new variables as well as the corresponding regressioncoefficients (i.e., neural network weights) that will be used as thenew encodings.

• step_woe() creates new variables based on weight of evidenceencodings.

• step_feature_hash() can create indicator variables using featurehashing.

For numeric predictors:

• step_umap() uses a nonlinear transformation similar to t-SNE but canbe used to project the transformation on new data. Both supervised andunsupervised methods can be used.

• step_discretize_xgb() andstep_discretize_cart() can make binnedversions of numeric predictors using supervised tree-based models.

• step_pca_sparse() andstep_pca_sparse_bayes() conduct featureextraction with sparsity of the component loadings.

Some references for these methods are:

• Francois C and Allaire JJ (2018)Deep Learning withR, Manning
• Guo, C and Berkhahn F (2016) “Entity Embeddings of CategoricalVariables”
• Micci-Barreca D (2001) “A preprocessing scheme for high-cardinalitycategorical attributes in classification and predictionproblems,”ACM SIGKDD Explorations Newsletter, 3(1), 27-32.
• Zumel N and Mount J (2017) “vtreat: adata.frame Processor forPredictive Modeling”
• McInnes L and Healy J (2018)UMAP: Uniform Manifold Approximation andProjection for Dimension Reduction
• Good, I. J. (1985), “Weight of evidence: A briefsurvey”,Bayesian Statistics, 2, pp.249-270.

There are two articles that walk through how to use these embeddingsteps, usinggeneralized linearmodels andneural networks built viaTensorFlow.

To install the package:

install.packages("embed")

Note that to use some steps, you will also have to install otherpackages such asrstanarm andlme4. For all of the steps to work,you may want to use:

install.packages(c("rpart","xgboost","rstanarm","lme4"))

To get a bug fix or to use a feature from the development version, youcan install the development version of this package from GitHub.

# install.packages("pak")pak::pak("tidymodels/embed")

This project is released with aContributor Code ofConduct.By contributing to this project, you agree to abide by its terms.

• For questions and discussions about tidymodels packages, modeling, andmachine learning, pleasepost on RStudioCommunity.

• If you think you have encountered a bug, pleasesubmit anissue.

• Either way, learn how to create and share areprex(a minimal, reproducible example), to clearly communicate about yourcode.

• Check out further details oncontributing guidelines for tidymodelspackages andhow to gethelp.