Super Learner Prediction
Implements the super learner prediction method and contains a
library of prediction algorithms to be used in the super learner.
News for the SuperLearner package.
- Added model.matrix to SL.xgboost
- Fixed innerCvControl in CV.SuperLearner to allow multiple parameters. It must now be a list of lists.
- create.Learner(): support character arguments.
- Glmnet: support alternative loss functions; when predicting automatically add any missing covariates and remove covariates not in the original data.
- Added SL.kernelKnn
- Added SL.ksvm
- Added SL.ranger
- Added vignette: "Guide to SuperLearner"
- Added SL.biglasso
- Added SL.lm, SL.speedlm, and SL.speedglm
- Added SL.lda and SL.qda
- Added SL.dbarts for C++-based bayesian additive regression trees.
- SL.lm and SL.glm now have a model argument, defaulting to TRUE (matching glm and lm), but can be changed to FALSE to conserve memory. Both wrappers also explicitly convert X matrix to a data frame.
- Added SL.extraTrees for extremely randomized trees, a random forest variant.
- Fixes prediction when a learner fails for methods: NNLS, NNloglik, CC_nloglik, and AUC. NNLS2 and CC_LS still have this bug. This fix required that an additional optional argument "errorsInLibrary" be passed to methods. This argument is a vector set to TRUE for learners that failed during model fitting.
- Add validRows option for CV.SuperLearner. Can now pass a cvControl for the outer CV and a list of cvControls, one for each cross-validation folds SuperLearner calls. default number of folds in CV.SuperLearner is now 10, matching the default with cvControl. If the user specifies both V and number of folds in cvControl(), an error message is returned.
- Added shrinkage parameter to SL.gbm
- fixed mtry default in SL.randomForest
- in CV.SuperLearner, fixed order for checking parallel options and folds argument in parLapply (thanks Chris Kennedy)
- updated method.AUC to change defaults on the optimization and add warnings for non-convergence
- Added wrapper for xgboost (thanks Chris Kennedy)
- Added wrapper for bartMachine (thanks Chris Kennedy)
- Added travis.ci checks
- Added environments for SuperLearner() and CV.SuperLearner() wrappers search path (includes SL., screen., and method.* wrappers)
- Added binary outcomes for SL.cforest
- Updated contact information
- Added additional svm() arguments for SL.svm
- Added recombineSL and recombineCVSL functions to re-fit the ensemble using a new metalearner in a computationally efficient manner
- For all wrappers, converted to format package::function when calling functions from other namespaces
- Added S3 method declarations for all predict.SL.* functions
- Added a
- Fixed error when computeCoef was re-run because of algorithms failing on full data
- Fixed Description field in Description file for CRAN policy
- Fixed check for method.AUC and family
- Moved SL.bart over to SuperLearneExtra because BayesTree package no longer on CRAN
- Added method.AUC, contributed by Erin LeDell
- added the SampleSplitSuperLearner function to allow sample split validation instead of V-fold cross-validation
- fixed package requirement in CV.SuperLearner from multicore to parallel
- Fixed a conflict with the reorder function in plot.CV.SuperLearner (between the stats and gdata namespace)
- Fixed a bug in SL.svm when family is binomial to grab the correct predicted probabilities (thanks to Jeremy Coyle)
- Added .Rbuildignore to not include the README.md file from GitHub on CRAN
- Removed SuperLearner.Rnw
- Moved vignettes to vignettes folder
- Changed cluster example to use PSOCK instead of MPI in SuperLearner.Rd
- removed the ":::" in plot.CV.SuperLearner
- moved quadprog from depends to suggests as it is only called if the user uses method = "method.NNLS2" not the default.
- Added method.CC_LS and method.CC_nloglik. These provide true convex combination optimization for the 2 loss functions. Contributed by Sam Lendle.
- Updated help documents
- Added links to SuperLearnerExtra on Github
- Switched from snow and multicore to parallel package
- fixed bug in CV.SuperLearner for leave-one-out cross-validation
- fixed bug in snowSuperLearner when only one screening algorithm is present
- method.NNloglik now reports the average -log likelihood instead of the sum to be consistent with NNLS
- Fixed bug in CV.SuperLearner not saving SuperLearner objects (watch out for ifelse() statements).
- Added minbucket to SL.rpart.
- Added SL.rpartPrune, a version of SL.rpart with built-in pruning.
- Minor changes to Rd files to cut build and check time. Time intensive examples now wrapped in \dontrun for CRAN.
- added plot.CV.SuperLearner
- fixed bug when one of the algorithms in SL.library has an error.
- fixed mcSuperLearner and snowSuperLearner not saving fitLibrary.
- added a placeholder Sweave vignette (SuperLearnerPresent.Rnw) to contain the SuperLearner presentation so the file can be found using the vignette() and browseVignettes() functions.
- CV.SuperLearner now outputs
- summary.CV.SuperLearner has returned
- added predict.SuperLearner
- Version 2.* represents a complete rewrite of the SuperLearner package.
- Details on the changes from Version 1.* to 2.* can be found in ChangeLog.