SuperLearnerMacro

Available Learners

1. Learners based on SAS procedures ( * denotes SAS Enterprise Miner/EM High Performance procedure)

   learner: description (SAS procedure)

   • ahal: approximate (faster) highly adaptive LASSO - uses non-parametric representation of data using indicator basis functions, selects LASSO parameter via minimum AICC across models using a range of 100 possible LASSO parameter values (HPGENSELECT)
   • ahalb: approximate (faster) highly adaptive LASSO, version 2 - uses non-parametric representation of data using indicator basis functions, selects LASSO parameter via 10-fold cross validation from a range of 100 possible LASSO parameter values (similar to the R function ‘hal’) - warning for binary outcome: does not respect [0,1] probability space (GLMSELECT)
   • back: backward selection by AICC (HPGENSELECT)
   • *bagging: bootstrap aggregation of regression/classification trees (ARBOR)
   • *bayesnet: Bayesian network [binary only] (HPBNET)
   • boxcox: Box-Cox transformation of target variables, positively bound continuous Y only (TRANSREG)
   • *boost: Gradient boosting of regression/classification trees (ARBOR)
   • cvcart: classification/regression tree with cross validated selection of final tree based on cost-complexity (HPSPLIT)
   • cart: classification/regression tree, no cross validation (HPSPLIT)
   • *deepnn: deep neural net with 5 hidden layers of 50 nodes each (DMDB/NEURAL)
   • enet: elastic net - warning for binary outcome: does not respect [0,1] probability space (GLMSELECT)
   • gam: generalized additive model with 3 df splines on all continuous variables (GAM)
   • gampl: faster generalized additive model with 3 df splines on all continuous variables (GAMPL)
   • glm: linear or logistic regression: slower wrappers for logit and linreg (GENMOD/REG)
   • hal: highly adaptive LASSO - uses non-parametric representation of data using indicator basis functions, selects LASSO parameter via 10-fold cross validation with exhaustive search - warning for binary outcome: does not respect [0,1] probability space (GLMSELECT)
   • hpglm: main term logistic/linear regression with multithreading (HPLOGISTIC/HPREG)
   • hplar: least angle regression with multithreading - warning for binary outcome: does not respect [0,1] probability space (HPREG)
   • hplogit: main term logistic regression with multithreading (HPLOGISTIC)
   • hplinreg: main term linear regression with multithreading (HPREG)
   • knn: [bernoulli only] k-nearest neighbors classification (DISCRIM)
   • lar: least angle regression - warning for binary outcome: does not respect [0,1] probability space (GLMSELECT)
   • lasso: least absolute shrinkage and selection operator with multithreading (HPGENSELECT)
   • lassob: LASSO with glmselect [use caution with binary variables] - may be appropriate for older sas versions (GLMSELECT)
   • cvlasso: LASSO with cross-validated selection of shrinkage parameter - warning for binary outcome: does not respect [0,1] probability space (GLMSELECT)
   • logit: main term logistic regression (GENMOD)
   • linreg: main term linear regression (REG)
   • mars: multivariate adaptive regression splines (ADAPTIVEREG)
   • mean: marginal mean of the prediction variable (GENMOD)
   • *nn: single hidden layer neural net with 5 nodes (HPNEURAL)
   • *nbayes: naive Bayes (HPBNET)
   • bspline: basis spline regression (TRANSREG)
   • pbspline: penalized basis spline regression [SINGLE CONTINUOUS PREDICTOR ONLY] (TRANSREG)
   • probit: main term probit regression (PROBIT)
   • ridge: ridge regression - warning for binary outcome: does not respect [0,1] probability space (REG)
   • quantreg: Quantile regression for the median - warning for binary outcome: does not respect [0,1] probability space (QUANTREG)
   • rf: random forest (HPFOREST)
   • rfoob: random forest, using out of bag predictions and modified selection criteria (HPFOREST)
   • *sherwood: random forest, sampling with replacement (ARBOR)
   • swise: Stepwise model selection - may be appropriate alternative to lasso for older sas versions (HPGENSELECT)
   • *svm: Support vector machine - linear kernel [binary only] (HPSVM)
   • *svmrbf: Support vector machine - radial basis function kernel [binary only] (HPSVM)

   Also includes multiple learners that are identical to other learners but include all first order interaction terms:

   backint, logitint, linregint, lassoint, lassobint, cvlassoint, swiseint, larint, enetint, gamint, gamplint, marsint, probitint, hpglmint, hplarint

2. Learners based on R functions

   A number of included R functions (requires SAS/IML and RLANG system option enabled (see here) allow a limited set of learners in the R programming language. Provided that R is installed and the RLANG option properly enabled, the required packages will be automatically installed the first time the learner is called (if running the SuperLearner or CVSuperLearner macros (internet connection must be active)). In most cases, the default parameterization of each of these learners is identical to the parameterization given in the SuperLearner package in R, so using them with the %SuperLearner macro should give similar results to using them in R. NOTE: R learners are case sensitive, unlike SAS, so specification of Y and predictor variable names must have correct case.

   • r_bagging: bootstrap aggregation of regression/classification trees (requires ipred package)
   • r_bart: Bayesian additive regression/classification trees (requires dbarts package)
   • r_boost: Gradient boosting regression/classification (requires xgboost package)
   • r_enet: elastic net regression/classification with cross validated selection of shrinkage parameter (requires glmnet package)
   • r_gam: generalized additive model (requires gam package)
   • r_lasso: LASSO regression/classification with cross validated selection of shrinkage parameter (requires glmnet package)
   • r_mars: MARS - multivariate adaptive regression splines (requires earth package)
   • r_polymars: MARS - multivariate adaptive polynomial regression splines (requires polspline package)
   • r_rf: random forest using R superlearner defaults (requires randomForest package)
   • r_ridge: ridge regression/classification with cross validated selection of shrinkage parameter (requires glmnet package)
   • r_rpart: CART - recursive partioning with no pruning (similar to cart) (requires rpart package)
   • r_rpartprune: CART - recursive partioning with pruning by complexity parameter selected through cross-validation (similar to cvcart) (requires rpart package)
   • r_sl: R based super learner (using 10-fold cross-validation, NNLS method) with several default learners (glm, lasso, bart, rf, mars; requires SuperLearner package)
   • r_svm: support vector machine regression/classification (requires e1071 package)

Main help page

Super learner macro home page