An Ensemble Method for Combining Subset-Specific Algorithm Fits

Subsemble is a general subset ensemble prediction method, which can be used for small, moderate, or large datasets. Subsemble partitions the full dataset into subsets of observations, fits a specified underlying algorithm on each subset, and uses a unique form of V-fold cross-validation to output a prediction function that combines the subset-specific fits. An oracle result provides a theoretical performance guarantee for Subsemble.


News

News for the subsemble package.

Notice to users: The interface (function arguments/values) may be subject to change prior to version 1.0.0.

Known Bugs

  • Cannot use parallel="multicore" when "SL.gbm" is part of the library. This is because the "SL.gbm" function uses multicore parallelization by default and R gets upset when you have two levels of multicore parallelism. This may be fixed in a future release. This is the error: Error in checkForRemoteErrors(val) : 7 nodes produced errors; first error: cannot open the connection
  • Some of the SuperLearner algorithm wrappers have issues with factor variables, for example SL.glm. This occurs in the internal CV step when the validation set has new levels that are not in the training set. These can be fixed by updating the wrapper functions.

To Do

  • Add support in the for the method function type from the SuperLearner::SuperLearner function as an option for the metalearner.
  • Add cvRisk to subsemble output. Currently NULL.
  • Add extra validation of learner and metalearner arguments using the .check.SL.library function from the SL package or write a validation function from scratch.
  • Drop learners that produce NA predicted values in Z matrix, or add some graceful handling of these cases.
  • Add the README.md file with an overview and examples.
  • Modify the parallel code to export only the specific objects required for computation instead of forking the entire environment. Maybe change the parallel backend to the doParallel package.
  • Add the Supervised Regression Tree subsemble option (activated by setting subControl$supervised = "SRT" or something similar).
  • Maybe disallow the number of subsets to be specified via the subsets argument and reserve that for specific subset (row index) lists only. It might make more sense to just set the number of subsets via the subControl argument.
  • Maybe add row.names to subpred data.frame.
  • Add a warning message when a user tries to predict using a subsemble fit that was saved using genControl$saveFits=FALSE. Currently, an unclear error message comes up: No applicable method for 'predict' applied to an object of class "NULL"
  • All parallel support for the subsemble.predict function.
  • Maybe use different seed arguments for data partitioning seed and model seed, so that we can set a partitioning seed, but leave the model seed as NULL. We could remove the seed argument and create subControl$seed, learnControl$seed list elements.
  • Modify the subsemble internals to accept wrappers using both the function(Y, X, newX) notation (from the SuperLearner package) and translate the functions to the function(x, y, newx) notation.
  • Allow user to specify the learner as a list instead of a vector.
  • Add support for screening algorithms in the learner list.

subsemble 0.0.9 (2014-07-01)

  • Re-named the argument x to xmat in subsemble internal functions .cvFun and .fitWrapper to avoid naming conflicts with parSapply functions. This was a bug introduced by the previous argument name change from X to x in the subsemble function.
  • Reduced the number of cores in a multicore cluster to the length of the list that the parSapply function is applied to (or number of cores, if that is fewer). Previously, the multicore cluster was unnecessarily created using the maximum number of cores available.

subsemble 0.0.8 (2014-06-28)

  • Changed the X, Y, newX arguments of the subsemble function to x, y, newx in order to conform to more common ML algorthim conventions in R. Additionally, the order of the first two arguments was changed from (Y,X) to (x, y) to match common convention. Lastly, the newdata argument in the predict.subsemble function was changed to newx. The predict function expects the new data.frame/matrix to have the same structure as the design matrix, x, hence it is more descriptive to name this argument newx.
  • Added a runtime element to the subsemble output which records the execution times of various steps in the algorithm.
  • Added startup messages in the R/zzz.R file.
  • Changed control.R functions to internal functions.

subsemble 0.0.7 (2014-06-26)

  • Added subsets = 1 functionality to train on full data (subsets = 1 is the traditional Super Learner algorithm), and also added an example of this to the documentation.
  • Fixed a bug caused by repeating the same wrapper function in the learner argument, ie. learner = c("SL.randomForest","SL.randomForest","SL.glm").
  • Fixed a typo in subsemble.Rd.
  • Changed the name of internal function .makeZ to .make_Z_l since it operates on a single learner only.
  • Replaced length(learner) with alias L in multiple places in the subsemble code.
  • Modified how names(Z) is specified.
  • Forcibly set stratifyCV = FALSE for both cvControl and subControl when the outcome is not binary.
  • git note: enable_j1_case branch merged into master branch.

subsemble 0.0.6 (2014-05-07)

  • Added the ability to pass a snow cluster object to the parallel argument of the subsemble function.
  • git note: snowcluster branch merged into master branch.

subsemble 0.0.5 (2014-05-05)

  • Modified predict.subsemble code to force individual learner predictions into a vector. The SL.earth wrapper created errors because it was returning a 1-column matrix instead of a vector. Also modified the function to assign column names the subpred data.frame after inserting the predictions instead of before inserting the predictions.
  • Added the genControl argument to the subsemble function to allow the user to not have to save all the model fits for a one-off type analysis. This also required adding a gen_control function to the control.R script and modifying some internal subsemble code.
  • Changed the parallel argument in the subsemble function from a logical argument to a string equal to "seq" (the default) or "multicore".
  • Inside the subsemble function, changed the name of the sub.cvlist object to subCVsets, which matches the output list name.
  • Change names of sublearner.predict and subsemble.predict output elements of the subsemble function to subpred and pred, respectively. Also updated the sublearner.predict output element of the predict.subsemble function to subpred. These were previously named to match the SuperLearner output conventions, but now they are nicer and easier to type.
  • Reversed the objects names sublearners and subfits inside the subsemble function because it made more sense to return an output object named subfits so it matches metafit. Also required updating the predict.subsemble function.
  • Removed internal function, .predFun since it was not being used for anything.
  • Updated internal variables L and l to M and m to be consistent with the notation in the technical paper.

subsemble 0.0.4 (2014-04-23)

  • Added an example to the subsemble function documentation. (Might need to update when the multiType option is added.)
  • Fixed bug in .subFun that happens when row.names(X) is not 1:N. Fixed by simply re-assigning row.names(X) as 1:N. Fixed: Error in FUN(1:3[[1L]], ...) : names not identical
  • Added the sub_control and learn_control functions and added a learnControl argument to the subsemble function. Also renamed the subsemble.CV.control function to cv_control.
  • Added multiType parameter to learn_control and implemented the "crossprod" type of library expansion.
  • Removed the shuffle parameter in subControl and added a supervision=NULL parameter. The latter is a placeholder for the Supervised Regression Tree subsemble, which is not implemented yet.
  • Removed validRows option in sub_control (previously, subsemble.CV.control), since it's not being used.
  • Updated argument descriptions in subsemble.Rd.
  • Fixed bug where seed was not being set. Also added set.seed(seed) inside functions (with inherent randomness, like training functions) that were being applied in parallel to ensure that the results would be the same regardless of whether the operations were being run in parallel or not.
  • Renamed internal function argument, x, to xmat to avoid conflict with internal clusterApply functions.
  • Added a .makeZ internal function in order to more easily switch between multiType modes.
  • Changed parLapply functions to parSapply in order to match output to sapply. This was causing the following error: "Error in cvRes[j, ] (from #145) : incorrect number of dimensions"
  • Removed validRows option in cv_control since it's not being used.
  • Switched the placement of the id argument in the subsemble function to later on in the argument list since it's not a commonly used argument.
  • Bug in predict.SL.knn when metalearner = "SL.knn": Error in as.matrix(train) : argument "X" is missing, with no default.
  • Added new list elements to the output of the subsemble function.

subsemble 0.0.3 (2014-03-29)

  • Fixed bug where predict(object, newdata) was not working for object of class "subsemble".
  • Fixed typo where the learner default argument was listed as "glm" instead of "SL.glm".
  • Added manual bound enforcement of Z matrix of predictions in the predict.subsemble function.
  • Added row names (same as input X row names) to the Z data.frame, a list element of the subsemble output.
  • Dropped require(SuperLearner) from the subsemble package since it is already loaded as a dependency.

subsemble 0.0.2 (2014-02-16)

  • Removed the dependency on SuperLearner.CV.control function in the SuperLearner package. Replaced with modified version of the function called, subsemble.CV.control. This updated version of the SuperLearner.CV.control function overrides the validRows argument (to be NULL).
  • Added support for multiple learners. Now the learner argument can be a character vector with a length that is a divisor of J.
  • Added support for the metalearner to be any general learner available in the SuperLearner candidate learner API, such as "SL.glmnet" or "SL.randomForest". Previously, the metalearner was just a glm.
  • Removed requirement that learner must be a string must be like "SL\\.[0-9a-zA_Z]+". Now validating the learner and metalearner arguments by using exists(learner).
  • Updated citation for Subsemble paper in the docs.
  • Fixed bug in .subFun in the subsemble function that caused errors when learners returned unlabeled predictions (missing row index name), as well as when they returned a 1-col data.frame instead of a vector for the pred object. The former occured with learner = c("SL.bart") and the latter occured with learner = c("SL.glmnet").

subsemble 0.0.1 (2014-01-19)

  • Initial release.

Reference manual

It appears you don't have a PDF plugin for this browser. You can click here to download the reference manual.

install.packages("subsemble")

0.0.9 by Erin LeDell, 4 years ago


Browse source code at https://github.com/cran/subsemble


Authors: Erin LeDell , Stephanie Sapp , Mark van der Laan


Documentation:   PDF Manual  


GPL (>= 2) license


Depends on SuperLearner

Suggests arm, caret, class, e1071, earth, gam, gbm, glmnet, Hmisc, ipred, lattice, LogicReg, MASS, mda, mlbench, nnet, parallel, party, polspline, quadprog, randomForest, rpart, SIS, spls, stepPlr


See at CRAN