In ModelOriented/DALEX: moDel Agnostic Language for Exploration and eXplanation

knitr::opts_chunk$set(
  collapse = FALSE,
  comment = "#>",
  warning = FALSE,
  message = FALSE
)

Intro

yardstick is a package that offers many measures for evaluating model performance. It is based on the tidymodels/tidyverse philosophy, the performance is calculated by functions working on the data.frame with the results of the model.

DALEX uses model performance measures to assess the importance of variables (in the model_parts function). These are typically calculated based on loss functions (functions with prefix loss) that are working on two vectors - the score from the model and the true target variable.

Although these packages have a slightly different philosophy of operation, you can use the measures available in yardstick when working with DALEX. Below is information on how to use the loss_yardstick function to do this.

Prepare a classification model

The yardstick package supports both classification models and regression models. We will start our example with a classification model for the titanic data - the probability of surviving this disaster.

The following instruction trains a classification model.

library("DALEX")
library("yardstick")
titanic_glm <- glm(survived~., data = titanic_imputed, family = "binomial")

Class Probability Metrics

The Class Probability Metrics in the yardstick package assume that the true value is a factor and the model returns a numerical score. So let's prepare an explainer that has factor as y and the predict_function returns the probability of the target class (default behaviour).

NOTE: Performance measures will be calculated on data supplied in the explainer. Put here the test data!

explainer_glm <- DALEX::explain(titanic_glm,
                        data = titanic_imputed[,-8],
                        y = factor(titanic_imputed$survived))

To make functions from the yardstick compatible with DALEX we must use the loss_yardstick adapter. In the example below we use the roc_auc function (area under the receiver operator curve). The yardstick:: prefix is not necessary, but we put it here to show explicitly where the functions you use are located.

NOTE: we set yardstick.event_first = FALSE as the model predicts probability of survived = 1.

options(yardstick.event_first = FALSE)

glm_auc <- model_parts(explainer_glm, type = "raw",
                  loss_function = loss_yardstick(yardstick::roc_auc))
glm_auc
plot(glm_auc)

In a similar way, we can use the pr_auc function (area under the precision recall curve).

glm_prauc <- model_parts(explainer_glm, type = "raw",
                  loss_function = loss_yardstick(yardstick::pr_auc))
glm_prauc
plot(glm_prauc)

Classification Metrics

The Classification Metrics in the yardstick package assume that the true value is a factor and the model returns a factor variable.

This is different behavior than for most explanations in DALEX, because when explaining predictions we typically operate on class membership probabilities. If we want to use Classification Metrics we need to provide a predict function that returns classes instead of probabilities.

So let's prepare an explainer that has factor as y and the predict_function returns classes.

explainer_glm <- DALEX::explain(titanic_glm,
                        data = titanic_imputed[,-8],
                        y = factor(titanic_imputed$survived),
                        predict_function = function(m,x) {
                              factor(as.numeric(predict(m, x, type = "response") > 0.5), 
                                     levels = c("0", "1"))
                        })

Again, let's use the loss_yardstick adapter. In the example below we use the accuracy function.

glm_accuracy <- model_parts(explainer_glm, type = "raw",
                    loss_function = loss_yardstick(yardstick::accuracy))
glm_accuracy
plot(glm_accuracy)

In a similar way, we can use the bal_accuracy function (balanced accuracy).

glm_bal_accuracy <- model_parts(explainer_glm, type = "raw",
                    loss_function = loss_yardstick(yardstick::bal_accuracy))
glm_bal_accuracy
plot(glm_bal_accuracy)

The lower the better?

For the loss function, the smaller the values the better the model. Therefore, the importance of variables is often calculated as loss(perturbed) - loss(original).

But many model performance functions have the opposite characteristic, the higher they are the better (e.g. AUC, accuracy, etc). To maintain a consistent analysis pipeline it is convenient to invert such functions, e.g. by converting to 1- AUC or 1 - accuracy.

To do it, just add the reverse = TRUE argument.

glm_1accuracy <- model_parts(explainer_glm, 
                    loss_function = loss_yardstick(accuracy, reverse = TRUE))
glm_1accuracy
plot(glm_1accuracy)

Calculate performance on whole dataset

By default the performance is calculated on N = 1000 randomly selected observations (to speed up the calculations). Set N = NULL to use the whole dataset.

glm_1accuracy <- model_parts(explainer_glm, 
                    loss_function = loss_yardstick(accuracy, reverse = TRUE),
                    N = NULL)
plot(glm_1accuracy)

Prepare a regression model

The following instruction trains a regression model.

library("ranger")
apartments_ranger <- ranger(m2.price~., data = apartments, num.trees = 50)

Regression Metrics

The Regression Metrics in the yardstick package assume that the true value is a numeric variable and the model returns a numeric score.

explainer_ranger  <- DALEX::explain(apartments_ranger, data = apartments[,-1],
                             y = apartments$m2.price, label = "Ranger Apartments")

To make functions from the yardstick compatible with DALEX we must use the loss_yardstick adapter. In the example below we use the rmse function (root mean squared error).

ranger_rmse <- model_parts(explainer_ranger, type = "raw",
                      loss_function = loss_yardstick(rmse))
ranger_rmse
plot(ranger_rmse)

And one more example for rsq function (R squared).

ranger_rsq <- model_parts(explainer_ranger, type = "raw",
                      loss_function = loss_yardstick(rsq))
ranger_rsq
plot(ranger_rsq)

Summary

I hope that using the yardstick package at DALEX will now be easy and enjoyable. If you would like to share your experience with this package, please create an issue at https://github.com/ModelOriented/DALEX/issues.