precrec (version 0.24.0)

prob_metrics: Calculate the Brier score, the RMSE and the log loss of predicted probabilities

Description

The prob_metrics function calculates three probability-based evaluation metrics - the Brier score, its square root the root mean squared error, and the log loss - for prediction scores that are probabilities. Unlike ROC and Precision-Recall curves, these metrics depend on the values of the scores rather than on their ranks, so the scores must lie in the range [0, 1].

Usage

prob_metrics(
  mdat,
  scores = NULL,
  labels = NULL,
  eps = 1e-15,
  metrics = NULL,
  ...
)

Value

The prob_metrics function returns a data frame with the columns modnames, dsids, metrics, and values. metrics is one of "brier", "rmse" or "logloss", so each model and dataset combination takes up three rows - plus one row for each metric named in metrics.

Arguments

mdat

An S3 object created by the mmdata() function. It contains formatted scores and labels. The prob_metrics function ignores scores and labels when mdat is specified. These arguments are internally passed to the mmdata() function when mdat is unspecified. In that case, both scores and labels must be at least specified.

scores

A numeric dataset of predicted probabilities. It can be a vector, a matrix, an array, a data frame, or a list.

labels

A numeric, character, logical, or factor dataset of observed labels. It can be a vector, a matrix, an array, a data frame, or a list.

eps

A numeric value used to clamp the scores away from 0 and 1 before the log loss is calculated. A single confident and wrong prediction would otherwise make the log loss infinite.

metrics

A character vector of additional metrics to calculate. The three metrics above are always returned; "d2_brier" and "d2_logloss" are returned as well when they are named here, and "all" asks for every metric the function knows.

A D2 score rescales a loss against the loss of the null model, the one that predicts the observed prevalence for every case and ignores the scores:

D2 = 1 - loss(model) / loss(null)

It is 1 for a perfect model and 0 for one that does no better than the prevalence, and it is negative for a model that does worse - which is a real result rather than an error, so it is not clipped. A dataset holding a single class has a null loss of 0 and so a D2 score of NA.

...

These additional arguments are passed to mmdata() for data preparation.

See Also

prob_metrics_ci() for the CIs of these metrics over multiple datasets. evalmod() for generating S3 objects with performance evaluation metrics. mmdata() for formatting input data.

Examples

##################################################
### Single model & single test dataset
###

## Predicted probabilities of 10 positives and 10 negatives
set.seed(1)
scores <- c(runif(10, 0.4, 1), runif(10, 0, 0.6))
labels <- c(rep(1, 10), rep(0, 10))

## Brier score and log loss
prob_metrics(scores = scores, labels = labels)


##################################################
### Multiple models & multiple test datasets
###

## The "poor_er" and "good_er" samples are drawn from beta distributions,
## so their scores can be read as probabilities
samps <- create_sim_samples(4, 100, 100, c("poor_er", "good_er"))
mdat <- mmdata(samps[["scores"]], samps[["labels"]],
  modnames = samps[["modnames"]],
  dsids = samps[["dsids"]]
)

pm <- prob_metrics(mdat)

## Show the Brier scores only
subset(pm, metrics == "brier")

## The D2 scores, which say how much of each loss the model explains
subset(
  prob_metrics(mdat, metrics = c("d2_brier", "d2_logloss")),
  dsids == 1
)