aum_diffs_penalty {aum}

aum diffs penalty

Description

Convert penalized errors to error differences. A typical use case is for penalized optimal changepoint models, for which small penalty values result in large fp/fn, and large penalty values result in small fp/fn.

Usage

aum_diffs_penalty(errors.df, 
    pred.name.vec, denominator = "count")

Arguments

errors.df

data.frame which describes error as a function of penalty/lambda, with at least columns example, min.lambda, fp, fn. Interpreted as follows: fp/fn occur from all penalties from min.lambda to the next value of min.lambda within the current value of example.

pred.name.vec

Character vector of prediction example names, used to convert names of label.vec to integers.

denominator

Type of diffs, either "count" or "rate".

Value

data table of class "aum_diffs" in which each rows represents a breakpoint in an error function. Columns are interpreted as follows: there is a change of "fp_diff","fn_diff" at predicted value "pred" for example/observation "example". This can be used for computing Area Under Minimum via aum function, and plotted via plot.aum_diffs.

Author(s)

Toby Dylan Hocking <[e-mail]> [aut, cre], Jadon Fowler [aut] (Contributed exact line search C++ code)

Examples

if(require("data.table"))setDTthreads(1L)#for CRAN check.

## Simple synthetic example with two changes in error function.
simple.df <- data.frame(
  example=1L,
  min.lambda=c(0, exp(1), exp(2), exp(3)),
  fp=c(6,2,2,0),
  fn=c(0,1,1,5))
(simple.diffs <- aum::aum_diffs_penalty(simple.df))
if(requireNamespace("ggplot2"))plot(simple.diffs)
(simple.rates <- aum::aum_diffs_penalty(simple.df, denominator="rate"))
if(requireNamespace("ggplot2"))plot(simple.rates)

## Simple real data with four example, one has non-monotonic fn.
if(requireNamespace("penaltyLearning")){
  data(neuroblastomaProcessed, package="penaltyLearning", envir=environment())
  ## assume min.lambda, max.lambda columns only? use names?
  nb.err <- with(neuroblastomaProcessed$errors, data.frame(
    example=paste0(profile.id, ".", chromosome),
    min.lambda,
    max.lambda,
    fp, fn))
  (nb.diffs <- aum::aum_diffs_penalty(nb.err, c("1.2", "1.1", "4.1", "4.2")))
  if(requireNamespace("ggplot2"))plot(nb.diffs)
}

## More complex real data example
data(fn.not.zero, package="aum", envir=environment())
pred.names <- unique(fn.not.zero$example)
(fn.not.zero.diffs <- aum::aum_diffs_penalty(fn.not.zero, pred.names))
if(requireNamespace("ggplot2"))plot(fn.not.zero.diffs)

if(require("ggplot2")){
  name2id <- structure(seq(0, length(pred.names)-1L), names=pred.names)
  fn.not.zero.wide <- fn.not.zero[, .(example=name2id[example], min.lambda, max.lambda, fp, fn)]
  fn.not.zero.tall <- data.table::melt(fn.not.zero.wide, measure=c("fp", "fn"))
  ggplot()+
    geom_segment(aes(
      -log(min.lambda), value,
      xend=-log(max.lambda), yend=value,
      color=variable, linewidth=variable),
      data=fn.not.zero.tall)+
    geom_point(aes(
      -log(min.lambda), value,
      fill=variable),
      color="black",
      shape=21,
      data=fn.not.zero.tall)+
    geom_vline(aes(
      xintercept=pred),
      data=fn.not.zero.diffs)+
    scale_size_manual(values=c(fp=2, fn=1))+
    facet_grid(example ~ ., labeller=label_both)
}