Grouping behavior in yardstick
In yardstick: Tidy Characterizations of Model Performance

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

The 1.3.0 release of yardstick introduced an implementation for groupwise metrics. The use case motivating the implementation of this functionality is fairness metrics, though groupwise metrics have applications beyond that domain. Fairness metrics quantify the degree of disparity in a metric value across groups. To learn more about carrying out fairness-oriented analyses with tidymodels, see the blog post on the tidymodels website. This vignette will instead focus on groupwise metrics generally, clarifying the meaning of "groupwise" and demonstrating functionality with an example dataset.

library(yardstick)
library(dplyr)

data("hpc_cv")

Group-awareness

Even before the implementation of groupwise metrics, all yardstick metrics had been group-aware. When grouped data is passed to a group-aware metric, it will return metric values calculated for each group.

To demonstrate, we'll make use of the hpc_cv data set, containing class probabilities and class predictions for a linear discriminant analysis fit to the HPC data set of Kuhn and Johnson (2013). The model is evaluated via 10-fold cross-validation, and the predictions for all folds are included.

tibble(hpc_cv)

For the purposes of this vignette, we'll also add a column batch to the data and select off the columns for the class probabilities, which we don't need.

set.seed(1)

hpc <-
  tibble(hpc_cv) %>%
  mutate(batch = sample(c("a", "b"), nrow(.), replace = TRUE)) %>%
  select(-c(VF, F, M, L))

hpc

If we wanted to compute the accuracy of the first resampled model, we could write:

hpc %>% 
  filter(Resample == "Fold01") %>%
  accuracy(obs, pred)

The metric function returns one row, giving the .metric, .estimator, and .estimate for the whole data set it is passed.

If we instead group the data by fold, metric functions like accuracy will know to compute values for each group; in the output, each row will correspond to a Resample.

hpc %>% 
  group_by(Resample) %>%
  accuracy(obs, pred)

Note that the first row, corresponding to Fold01, gives the same value as manually filtering for the observations corresponding to the first resample and then computing the accuracy.

This behavior is what we mean by group-awareness. When grouped data is passed to group-aware metric functions, they will return values for each group.

Groupwise metrics

Groupwise metrics are associated with a data-column such that, when passed data with that column, the metric will temporarily group by that column, compute values for each of the groups defined by the column, and then aggregate the values computed for the temporary grouping back to the level of the input data's grouping.

More concretely, let's turn to an example where there is no pre-existing grouping in the data. Consider the portion of the HPC data pertaining to the first resample:

hpc %>% 
  filter(Resample == "Fold01")

Suppose that the batches in the data represent two groups for which model performance ought not to differ. To quantify the degree to which model performance differs for these two groups, we could compute accuracy values for either group separately, and then take their difference. First, computing accuracies:

acc_by_group <- 
  hpc %>% 
  filter(Resample == "Fold01") %>%
  group_by(batch) %>%
  accuracy(obs, pred)

acc_by_group

Now, taking the difference:

diff(c(acc_by_group$.estimate[2], acc_by_group$.estimate[1]))

Groupwise metrics encode the group_by() and aggregation step (in this case, subtraction) shown above into a yardstick metric. We can define a new groupwise metric with the new_groupwise_metric() function:

accuracy_diff <-
  new_groupwise_metric(
    fn = accuracy,
    name = "accuracy_diff",
    aggregate = function(acc_by_group) {
      diff(c(acc_by_group$.estimate[2], acc_by_group$.estimate[1]))
    }
  )

The fn argument is the yardstick metric that will be computed for each data group.
The name argument gives the name of the new metric we've created; we'll call ours "accuracy difference."
The aggregate argument is a function defining how to go from fn output by group to a single numeric value.

The output, accuracy_diff, is a function subclass called a metric_factory:

class(accuracy_diff)

accuracy_diff now knows to take accuracy values for each group and then return the difference between the accuracy for the first and second result as output. The last thing we need to associate with the object is the name of the grouping variable to pass to group_by(); we can pass that variable name to accuracy_diff to do so:

accuracy_diff_by_batch <- accuracy_diff(batch)

The output, accuracy_diff_by_batch, is a yardstick metric function like any other:

class(accuracy)

class(accuracy_diff_by_batch)

We can use the accuracy_diff_by_batch() metric in the same way that we would use accuracy(). On its own:

hpc %>% 
  filter(Resample == "Fold01") %>%
  accuracy_diff_by_batch(obs, pred)

We can also add accuracy_diff_by_batch() to metric sets:

acc_ms <- metric_set(accuracy, accuracy_diff_by_batch)

hpc %>% 
  filter(Resample == "Fold01") %>%
  acc_ms(truth = obs, estimate = pred)

Groupwise metrics are group-aware. When passed data with any grouping variables other than the column passed as the first argument to accuracy_diff()---in this case, group---accuracy_diff_by_batch() will behave like any other yardstick metric. For example:

hpc %>% 
  group_by(Resample) %>%
  accuracy_diff_by_batch(obs, pred)

Groupwise metrics form the backend of fairness metrics in tidymodels. To learn more about groupwise metrics and their applications in fairness problems, see new_groupwise_metric().