tune: Hyperparameter tuning for classifiers
In CMA: Synthesis of microarray-based classification

Description Usage Arguments Details Value Note Author(s) References See Also Examples

Most classifiers implemented in this package depend on one or even several hyperparameters (s. details) that should be optimized to obtain good (and comparable !) results. As tuning scheme, we propose three fold Cross-Validation on each learningset (for fixed selected variables). Note that learningsets usually do not contain the complete dataset, so tuning involves a second level of splitting the dataset. Increasing the number of folds leads to larger datasets (and possibly to higher accuracy), but also to higher computing times.
For S4 method information, s. link{tune-methods}

1	tune(X, y, f, learningsets, genesel, genesellist = list(), nbgene, classifier, fold = 3, strat = FALSE, grids = list(), trace = TRUE, ...)

`X`	Gene expression data. Can be one of the following: A `matrix`. Rows correspond to observations, columns to variables. A `data.frame`, when `f` is not missing (s. below). An object of class `ExpressionSet`.
`y`	Class labels. Can be one of the following: A `numeric` vector. A `factor`. A `character` if `X` is an `ExpressionSet` that specifies the phenotype variable. `missing`, if `X` is a `data.frame` and a proper formula `f` is provided.
`f`	A two-sided formula, if `X` is a `data.frame`. The left part correspond to class labels, the right to variables.
`learningsets`	An object of class `learningsets`. May be missing, then the complete datasets is used as learning set.
`genesel`	Optional (but usually recommended) object of class `genesel` containing variable importance information for the argument `learningsets`
`genesellist`	In the case that the argument `genesel` is missing, this is an argument list passed to `GeneSelection`. If both `genesel` and `genesellist` are missing, no variable selection is performed.
`nbgene`	Number of best genes to be kept for classification, based on either `genesel` or the call to `GeneSelection` using `genesellist`. In the case that both are missing, this argument is not necessary. note: If the gene selection method has been one of `"lasso", "elasticnet", "boosting"`, `nbgene` will be reset to `min(s, nbgene)` where `s` is the number of nonzero coefficients. if the gene selection scheme has been `"one-vs-all", "pairwise"` for the multiclass case, there exist several rankings. The top `nbgene` will be kept of each of them, so the number of effective used genes will sometimes be much larger.
`classifier`	Name of function ending with `CMA` indicating the classifier to be used.
`fold`	The number of cross-validation folds used within each `learningset`. Default is 3. Increasing `fold` will lead to higher computing times.
`strat`	Should stratified cross-validation according to the class proportions in the complete dataset be used ? Default is `FALSE`.
`grids`	A named list. The names correspond to the arguments to be tuned, e.g. `k` (the number of nearest neighbours) for `knnCMA`, or `cost` for `svmCMA`. Each element is a `numeric` vector defining the grid of candidate values. Of course, several hyperparameters can be tuned simultaneously (though requiring much time). By default, `grids` is an empty list. In that case, a pre-defined list will be used, s. details.
`trace`	Should progress be traced ? Default is `TRUE`.
`...`	Further arguments to be passed to `classifier`, of course not one of the arguments to be tuned (!).

The following default settings are used, if the arguments grids is an empty list:

gbmCMA: n.trees = c(50, 100, 200, 500, 1000)
compBoostCMA: mstop = c(50, 100, 200, 500, 1000)
LassoCMA: norm.fraction = seq(from=0.1, to=0.9, length=9)
ElasticNetCMA: norm.fraction = seq(from=0.1, to=0.9, length=5), alpha = 2^{-(5:1)}
plrCMA: lambda = 2^{-4:4}
pls_ldaCMA: comp = 1:10
pls_lrCMA: comp = 1:10
pls_rfCMA: comp = 1:10
rfCMA: mtry = ceiling(c(0.1, 0.25, 0.5, 1, 2)*sqrt(ncol(X))), nodesize = c(1,2,3)
knnCMA: k=1:10
pknnCMA: k = 1:10
scdaCMA: delta = c(0.1, 0.25, 0.5, 1, 2, 5)
pnnCMA: sigma = c(2^{-2:2})

nnetCMA: size = 1:5, decay = c(0, 2^{-(4:1)})
svmCMA, kernel = "linear": cost = c(0.1, 1, 5, 10, 50, 100, 500)
svmCMA, kernel = "radial": cost = c(0.1, 1, 5, 10, 50, 100, 500), gamma = 2^{-2:2}
svmCMA, kernel = "polynomial": cost = c(0.1, 1, 5, 10, 50, 100, 500), degree = 2:4

An object of class tuningresult

The computation time can be enormously high. Note that for each different learningset, the classifier must be trained fold times number of possible different hyperparameter combinations times. E.g. if the number of the learningsets is fifty, fold = 3 and two hyperparameters (each with 5 candidate values) are tuned, 50x3x25=3750 training iterations are necessary !

Martin Slawski ms@cs.uni-sb.de

Anne-Laure Boulesteix boulesteix@ibe.med.uni-muenchen.de

Christoph Bernau bernau@ibe.med.uni-muenchen.de

Slawski, M. Daumer, M. Boulesteix, A.-L. (2008) CMA - A comprehensive Bioconductor package for supervised classification with high dimensional data. BMC Bioinformatics 9: 439

tuningresult, GeneSelection, classification

## Not run: 
### simple example for a one-dimensional grid, using compBoostCMA.
### dataset
data(golub)
golubY <- golub[,1]
golubX <- as.matrix(golub[,-1])
### learningsets
set.seed(111)
lset <- GenerateLearningsets(y=golubY, method = "CV", fold=5, strat =TRUE)
### tuning after gene selection with the t.test
tuneres <- tune(X = golubX, y = golubY, learningsets = lset,
              genesellist = list(method = "t.test"),
              classifier=compBoostCMA, nbgene = 100,
              grids = list(mstop = c(50, 100, 250, 500, 1000)))
### inspect results
show(tuneres)
best(tuneres)
plot(tuneres, iter = 3)

## End(Not run)