Classification parameter optimisation for the PerTurbo algorithm

perTurboOptimisation(
  object,
  fcol = "markers",
  pRegul = 10^(seq(from = -1, to = 0, by = 0.2)),
  sigma = 10^(seq(from = -1, to = 1, by = 0.5)),
  inv = c("Inversion Cholesky", "Moore Penrose", "solve", "svd"),
  reg = c("tikhonov", "none", "trunc"),
  times = 1,
  test.size = 0.2,
  xval = 5,
  fun = mean,
  seed,
  verbose = TRUE
)

Arguments

object

An instance of class "MSnSet".

fcol

The feature meta-data containing marker definitions. Default is markers.

pRegul

The hyper-parameter for the regularisation (values are in ]0,1] ). If reg =="trunc", pRegul is for the percentage of eigen values in matrix. If reg =="tikhonov", then 'pRegul' is the parameter for the tikhonov regularisation. Available configurations are : "Inversion Cholesky" - ("tikhonov" / "none"), "Moore Penrose" - ("tikhonov" / "none"), "solve" - ("tikhonov" / "none"), "svd" - ("tikhonov" / "none" / "trunc").

sigma

The hyper-parameter.

inv

The type of algorithm used to invert the matrix. Values are : "Inversion Cholesky" (chol2inv), "Moore Penrose" (ginv), "solve" (solve), "svd" (svd). Default value is "Inversion Cholesky".

reg

The type of regularisation of matrix. Values are "none", "trunc" or "tikhonov". Default value is "tikhonov".

times

The number of times internal cross-validation is performed. Default is 100.

test.size

The size of test data. Default is 0.2 (20 percent).

xval

The n-cross validation. Default is 5.

fun

The function used to summarise the times macro F1 matrices.

seed

The optional random number generator seed.

verbose

A logical defining whether a progress bar is displayed.

Value

An instance of class "GenRegRes".

Details

Note that when performance scores precision, recall and (macro) F1 are calculated, any NA values are replaced by 0. This decision is motivated by the fact that any class that would have either a NA precision or recall would result in an NA F1 score and, eventually, a NA macro F1 (i.e. mean(F1)). Replacing NAs by 0s leads to F1 values of 0 and a reduced yet defined final macro F1 score.

See also

perTurboClassification and example therein.

Author

Thomas Burger and Samuel Wieczorek