labelError {penaltyLearning}

Compute incorrect labels

Description

Compute incorrect labels for several change-point detection problems and models. Use this function after having computed changepoints, loss values, and model selection functions (see modelSelection). The next step after labelError is typically computing target intervals of log(penalty) values that predict changepoints with minimum incorrect labels for each problem (see targetIntervals).

Usage

labelError(models, labels, 
    changes, change.var = "chromStart", 
    label.vars = c("min", 
        "max"), model.vars = "n.segments", 
    problem.vars = character(0), 
    annotations = change.labels)

Arguments

models

data.frame with one row per (problem,model) combination, typically the output of modelSelection(...). There is a row for each changepoint model that could be selected for a particular segmentation problem. There should be columns problem.vars (for problem ID) and model.vars (for model complexity).

labels

data.frame with one row per (problem,region). Each label defines a region in a particular segmentation problem, and a range of predicted changepoints which are consistent in that region. There should be a column "annotation" with takes one of the corresponding values in the annotation column of change.labels (used to determine the range of predicted changepoints which are consistent). There should also be a columns problem.vars (for problem ID) and label.vars (for region start/end).

changes

data.frame with one row per (problem,model,change), for each predicted changepoint (in each model and segmentation problem). Should have columns problem.vars (for problem ID), model.vars (for model complexity), and change.var (for changepoint position).

change.var

character(length=1): column name of predicted change-point position in labels. The default "chromStart" is useful for genomic data with segment start/end positions stored in columns named chromStart/chromEnd. A predicted changepoint at position X is interpreted to mean a changepoint between X and X+1.

label.vars

character(length=2): column names of start and end positions of labels, in same units as change-point positions. The default is c("min", "max"). Labeled regions are (start,end] – open on the left and closed on the right, so for example a 0changes annotation between start=10 and end=20 means that any predicted changepoint at 11, ..., 20 is a false positive.

model.vars

character: column names used to identify model complexity. The default "n.segments" is for change-point models such as in the jointseg and changepoint packages.

problem.vars

character: column names used to identify data set / segmentation problem, should be present in all three data tables (models, labels, changes).

annotations

data.table with columns annotation, min.changes, max.changes, possible.fn, possible.fp which is joined to labels in order to determine how to compute false positives and false negatives for each annotation.

Value

list of two data.tables: label.errors has one row for every combination of models and labels, with status column that indicates whether or not that model commits an error in that particular label; model.errors has one row per model, with columns for computing target intervals and ROC curves (see targetIntervals and ROChange).

Author(s)

Toby Dylan Hocking

Examples

label <- function(annotation, min, max){
  data.frame(profile.id=4, chrom="chr14", min, max, annotation)
}
label.df <- rbind(
  label("1change", 70e6, 80e6),
  label("0changes", 20e6, 60e6))
model.df <- data.frame(chrom="chr14", n.segments=1:3)
change.df <- data.frame(chrom="chr14", rbind(
  data.frame(n.segments=2, changepoint=75e6),
  data.frame(n.segments=3, changepoint=c(75e6, 50e6))))
penaltyLearning::labelError(
  model.df, label.df, change.df,
  problem.vars="chrom", # for all three data sets.
  model.vars="n.segments", # for changes and selection.
  change.var="changepoint", # column of changes with breakpoint position.
  label.vars=c("min", "max")) # limit of labels in ann.