SLmetrics (version 0.3-4)

pr.curve: Precision Recall Curve

Description

A generic S3 function to compute the precision recall curve score for a classification model. This function dispatches to S3 methods in pr.curve() and performs no input validation. If you supply NA values or vectors of unequal length (e.g. length(x) != length(y)), the underlying C++ code may trigger undefined behavior and crash your R session.

Defensive measures

Because pr.curve() operates on raw pointers, pointer-level faults (e.g. from NA or mismatched length) occur before any R-level error handling. Wrapping calls in try() or tryCatch() will not prevent R-session crashes.

To guard against this, wrap pr.curve() in a "safe" validator that checks for NA values and matching length, for example:

safe_pr.curve <- function(x, y, ...) {
  stopifnot(
    !anyNA(x), !anyNA(y),
    length(x) == length(y)
  )
  pr.curve(x, y, ...)
}

Apply the same pattern to any custom metric functions to ensure input sanity before calling the underlying C++ code.

Area under the curve

Use auc.pr.curve for calculating the area under the curve directly.

Efficient multi-metric evaluation

To avoid sorting the same probability matrix multiple times (once per class or curve), you can precompute a single set of sort indices and pass it via the indices argument. This reduces the overall cost from O(K·N log N) to O(N log N + K·N).

## presort response
## probabilities
indices <- preorder(response, decreasing = TRUE)
## evaluate precision recall curve
pr.curve(actual, response, indices = indices)

Usage

## Generic S3 method
## for Precision Recall Curve
pr.curve(...)
## Generic S3 method
## for weighted Precision Recall Curve
weighted.pr.curve(...)

Value

A data.frame on the following form,

threshold

<numeric> Thresholds used to determine recall() and precision()

level

<character> The level of the actual <factor>

label

<character> The levels of the actual <factor>

recall

<numeric> The recall

precision

<numeric> The precision

Arguments

...

Arguments passed on to pr.curve.factor, weighted.pr.curve.factor

actual

A vector length \(n\), and \(k\) levels. Can be of integer or factor.

response

A \(n \times k\) <double>-matrix of predicted probabilities. The \(i\)-th row should sum to 1 (i.e., a valid probability distribution over the \(k\) classes). The first column corresponds to the first factor level in actual, the second column to the second factor level, and so on.

indices

An optional \(n \times k\) matrix of <integer> values of sorted response probability indices.

thresholds

An optional <double> vector of length \(n\) (default: NULL).

w

A <double> vector of sample weights.

References

James, Gareth, et al. An introduction to statistical learning. Vol. 112. No. 1. New York: springer, 2013.

Hastie, Trevor. "The elements of statistical learning: data mining, inference, and prediction." (2009).

Pedregosa, Fabian, et al. "Scikit-learn: Machine learning in Python." the Journal of machine Learning research 12 (2011): 2825-2830.

See Also

Other Classification: accuracy(), auc.pr.curve(), auc.roc.curve(), baccuracy(), brier.score(), ckappa(), cmatrix(), cross.entropy(), dor(), fbeta(), fdr(), fer(), fmi(), fpr(), hammingloss(), jaccard(), logloss(), mcc(), nlr(), npv(), plr(), precision(), recall(), relative.entropy(), roc.curve(), shannon.entropy(), specificity(), zerooneloss()

Other Supervised Learning: accuracy(), auc.pr.curve(), auc.roc.curve(), baccuracy(), brier.score(), ccc(), ckappa(), cmatrix(), cross.entropy(), deviance.gamma(), deviance.poisson(), deviance.tweedie(), dor(), fbeta(), fdr(), fer(), fmi(), fpr(), gmse(), hammingloss(), huberloss(), jaccard(), logloss(), maape(), mae(), mape(), mcc(), mpe(), mse(), nlr(), npv(), pinball(), plr(), precision(), rae(), recall(), relative.entropy(), rmse(), rmsle(), roc.curve(), rrmse(), rrse(), rsq(), shannon.entropy(), smape(), specificity(), zerooneloss()

Examples

## Classes and
## seed
set.seed(1903)
classes <- c("Kebab", "Falafel")

## Generate actual classes
## and response probabilities
actual_classes <- factor(
    x = sample(
      x = classes, 
      size = 1e2, 
      replace = TRUE, 
      prob = c(0.7, 0.3)
    )
)

response_probabilities <- ifelse(
    actual_classes == "Kebab", 
    rbeta(sum(actual_classes == "Kebab"), 2, 5), 
    rbeta(sum(actual_classes == "Falafel"), 5, 2)
)

## Construct response
## matrix
probability_matrix <- cbind(
    response_probabilities,
    1 - response_probabilities
)

## Visualize precision recall curve

plot(
    SLmetrics::pr.curve(
     actual   = actual_classes, 
     response = probability_matrix
 )
)