Postprocessors refine predictions outputted from machine learning models to improve predictive performance or better satisfy distributional limitations. This package introduces ‘tailor’ objects, which compose iterative adjustments to model predictions. In addition to utilities to create new adjustments, the package provides a number of pre-written ones:
- For probability distributions: calibration
- For transformation of probabilities to hard class predictions: thresholds, equivocal zones
- For numeric distributions: calibration, range
Tailors are tightly integrated with the tidymodels framework. For greatest ease of use, situate tailors in model workflows with add_tailor().
The package is under active development; please treat it as experimental and don’t depend on the syntax staying the same.
Installation
You can install the development version of tailor like so:
pak::pak("tidymodels/tailor")Example
The two_class_example dataset from modeldata gives the true value of an outcome variable truth as well as predicted probabilities (Class1 and Class2). The hard class predictions, in predicted, are "Class1" if the probability assigned to "Class1" is above .5, and "Class2" otherwise.
library(modeldata)
library(dplyr)
#>
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#>
#> filter, lag
#> The following objects are masked from 'package:base':
#>
#> intersect, setdiff, setequal, union
head(two_class_example)
#> truth Class1 Class2 predicted
#> 1 Class2 0.003589243 0.9964107574 Class2
#> 2 Class1 0.678621054 0.3213789460 Class1
#> 3 Class2 0.110893522 0.8891064779 Class2
#> 4 Class1 0.735161703 0.2648382969 Class1
#> 5 Class2 0.016239960 0.9837600397 Class2
#> 6 Class1 0.999275071 0.0007249286 Class1The model predicts "Class1" more often than it does "Class2".
If we wanted the model to predict "Class2" more often, we could increase the probability threshold assigned to "Class1" above which the hard class prediction will be "Class1". In the tailor package, this adjustment is implemented in adjust_probability_threshold(), which can be situated in a tailor object.
post_obj <-
tailor() %>%
adjust_probability_threshold(threshold = .9)
post_obj
#>
#> ── tailor ──────────────────────────────────────────────────────────────────────
#> A postprocessor with 1 adjustment:
#>
#> • Adjust probability threshold to 0.9.tailors must be fitted before they can predict on new data. For adjustments like adjust_probability_threshold(), there’s no training that actually happens at the fit() step besides recording the name and type of relevant variables. For other adjustments, like probability calibration with adjust_probability_calibration(), parameters are actually estimated at the fit() step and separate data should be used to train the postprocessor and evaluate its performance.
In this case, though, we can fit() on the whole dataset. The resulting object is still a tailor, but is now flagged as trained.
post_res <- fit(
post_obj,
two_class_example,
outcome = c(truth),
estimate = c(predicted),
probabilities = c(Class1, Class2)
)
post_res
#>
#> ── tailor ──────────────────────────────────────────────────────────────────────
#> A binary postprocessor with 1 adjustment:
#>
#> • Adjust probability threshold to 0.9. [trained]When used with a model workflow via add_tailor(), the arguments to fit() a tailor will be set automatically (in addition to the data splitting needed for postprocessors that require training).
Now, when passed new data, the trained tailor will determine the outputted class based on whether the probability assigned to the level "Class1" is above .9, resulting in more predictions of "Class2" than before.
predict(post_res, two_class_example) %>% count(predicted)
#> # A tibble: 2 × 2
#> predicted n
#> <fct> <int>
#> 1 Class1 180
#> 2 Class2 320Tailors compose adjustments; when several adjust_*() functions are called iteratively, tailors will apply them in order at fit() and predict() time.