IntervalRegressionCV {penaltyLearning}

IntervalRegressionCV

Description

Use cross-validation to fit an L1-regularized linear interval regression model by optimizing margin and/or regularization parameters. This function repeatedly calls IntervalRegressionRegularized, and by default assumes that margin=1. To optimize the margin, specify the margin.vec parameter manually, or use IntervalRegressionCVmargin (which takes more computation time but yields more accurate models). If the future package is available, two levels of future_lapply are used to parallelize on validation.fold and margin.

Usage

IntervalRegressionCV(feature.mat, 
    target.mat, n.folds = ifelse(nrow(feature.mat) < 
        10, 3L, 5L), 
    fold.vec = sample(rep(1:n.folds, 
        l = nrow(feature.mat))), 
    verbose = 0, min.observations = 10, 
    reg.type = "min", 
    incorrect.labels.db = NULL, 
    initial.regularization = 0.001, 
    margin.vec = 1, LAPPLY = NULL, 
    check.unlogged = TRUE, 
    ...)

Arguments

feature.mat

Numeric feature matrix, n observations x p features.

target.mat

Numeric target matrix, n observations x 2 limits. These should be real-valued (possibly negative). If your data are interval censored positive-valued survival times, you need to log them to obtain target.mat.

n.folds

Number of cross-validation folds.

fold.vec

Integer vector of fold id numbers.

verbose

numeric: 0 for silent, bigger numbers (1 or 2) for more output.

min.observations

stop with an error if there are fewer than this many observations.

reg.type

Either "1sd" or "min" which specifies how the regularization parameter is chosen during the internal cross-validation loop. min: first take the mean of the K-CV error functions, then minimize it (this is the default since it tends to yield the least test error). 1sd: take the most regularized model with the same margin which is within one standard deviation of that minimum (this model is typically a bit less accurate, but much less complex, so better if you want to interpret the coefficients).

incorrect.labels.db

either NULL or a data.table, which specifies the error function to compute for selecting the regularization parameter on the validation set. NULL means to minimize the squared hinge loss, which measures how far the predicted log(penalty) values are from the target intervals. If a data.table is specified, its first key should correspond to the rownames of feature.mat, and columns min.log.lambda, max.log.lambda, fp, fn, possible.fp, possible.fn; these will be used with ROChange to compute the AUC for each regularization parameter, and the maximimum will be selected (in the plot this is negative.auc, which is minimized). This data.table can be computed via labelError(modelSelection(...),...)$model.errors – see example(ROChange). In practice this makes the computation longer, and it should only result in more accurate models if there are many labels per data sequence.

initial.regularization

Passed to IntervalRegressionRegularized.

margin.vec

numeric vector of margin size hyper-parameters. The computation time is linear in the number of elements of margin.vec – more values takes more computation time, but yields slightly more accurate models (if there is enough data).

LAPPLY

Function to use for parallelization, by default future_lapply if it is available, otherwise lapply. For debugging with verbose>0 it is useful to specify LAPPLY=lapply in order to interactively see messages, before all parallel processes end.

check.unlogged

If TRUE, stop with an error if target matrix is non-negative and has any big difference in successive quantiles (this is an indicator that the user probably forgot to log their outputs).

...

passed to IntervalRegressionRegularized.

Value

List representing regularized linear model.

Author(s)

Toby Dylan Hocking

Examples

if(interactive()){
  library(penaltyLearning)
  data("neuroblastomaProcessed", package="penaltyLearning", envir=environment())
  if(require(future)){
    plan(multiprocess)
  }
  set.seed(1)
  i.train <- 1:100
  fit <- with(neuroblastomaProcessed, IntervalRegressionCV(
    feature.mat[i.train,], target.mat[i.train,],
    verbose=0))
  ## When only features and target matrices are specified for
  ## training, the squared hinge loss is used as the metric to
  ## minimize on the validation set.
  plot(fit)
  ## Create an incorrect labels data.table (first key is same as
  ## rownames of feature.mat and target.mat).
  library(data.table)
  errors.per.model <- data.table(neuroblastomaProcessed$errors)
  errors.per.model[, pid.chr := paste0(profile.id, ".", chromosome)]
  setkey(errors.per.model, pid.chr)
  set.seed(1)
  fit <- with(neuroblastomaProcessed, IntervalRegressionCV(
    feature.mat[i.train,], target.mat[i.train,],
    ## The incorrect.labels.db argument is optional, but can be used if
    ## you want to use AUC as the CV model selection criterion.
    incorrect.labels.db=errors.per.model))
  plot(fit)
}