penaltyLearning (version 2020.5.13)

targetIntervalROC: targetIntervalROC

Description

Compute a ROC curve using a target interval matrix. A prediction less than the lower limit is considered a false positive (penalty too small, too many changes), and a prediction greater than the upper limit is a false negative (penalty too large, too few changes). WARNING: this ROC curve is less detailed than the one you get from ROChange! Use ROChange if possible.

Usage

targetIntervalROC(target.mat, 
    pred)

Arguments

target.mat

n x 2 numeric matrix: target intervals of log(penalty) values that yield minimal incorrect labels.

pred

numeric vector: predicted log(penalty) values.

Value

list describing ROC curves, same as ROChange.

Examples

# NOT RUN {
library(penaltyLearning)
library(data.table)
data(neuroblastomaProcessed, envir=environment())

pid.vec <- c("1", "4")
chr <- 2
incorrect.labels <-
  neuroblastomaProcessed$errors[profile.id%in%pid.vec & chromosome==chr]
pid.chr <- paste0(pid.vec, ".", chr)
target.mat <- neuroblastomaProcessed$target.mat[pid.chr, , drop=FALSE]
pred.dt <- data.table(profile.id=pid.vec, pred.log.lambda=1.5)
roc.list <- list(
  labels=ROChange(incorrect.labels, pred.dt, "profile.id"),
  targets=targetIntervalROC(target.mat, pred.dt$pred.log.lambda))

err <- data.table(incorrect=names(roc.list))[, {
  roc.list[[incorrect]]$roc
}, by=incorrect]
library(ggplot2)
ggplot()+
  ggtitle("incorrect targets is an approximation of incorrect labels")+
  scale_size_manual(values=c(labels=2, targets=1))+
  geom_segment(aes(
    min.thresh, errors,
    color=incorrect,
    size=incorrect,
    xend=max.thresh, yend=errors),
               data=err)

# }