How to use DALEX with the yardstick package?
Przemyslaw Biecek
2022-09-07
Source:vignettes/vignette_yardstick.Rmd
vignette_yardstick.RmdIntro
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$survived))To make functions from the yardstick compatible with DALEX we must use the loss_yardstick adapter. In the example below we use the roc_auc function (area under the receiver operator curve). The yardstick:: prefix is not necessary, but we put it here to show explicitly where the functions you use are located.
NOTE: we set yardstick.event_first = FALSE as the model predicts probability of survived = 1.
options(yardstick.event_first = FALSE)
glm_auc <- model_parts(explainer_glm, type = "raw",
loss_function = loss_yardstick(yardstick::roc_auc))
glm_auc#> variable mean_dropout_loss label
#> 1 _full_model_ 0.8137302 lm
#> 2 gender 0.6174230 lm
#> 3 class 0.7166700 lm
#> 4 age 0.7863548 lm
#> 5 sibsp 0.8033765 lm
#> 6 embarked 0.8074437 lm
#> 7 fare 0.8133961 lm
#> 8 parch 0.8137905 lm
#> 9 _baseline_ 0.4871857 lm
plot(glm_auc)
In a similar way, we can use the pr_auc function (area under the precision recall curve).
glm_prauc <- model_parts(explainer_glm, type = "raw",
loss_function = loss_yardstick(yardstick::pr_auc))
glm_prauc#> variable mean_dropout_loss label
#> 1 _full_model_ 0.7415507 lm
#> 2 gender 0.4459565 lm
#> 3 class 0.5865294 lm
#> 4 age 0.7162901 lm
#> 5 embarked 0.7273457 lm
#> 6 sibsp 0.7277256 lm
#> 7 parch 0.7416492 lm
#> 8 fare 0.7418198 lm
#> 9 _baseline_ 0.3273634 lm
plot(glm_prauc)
Classification Metrics
The Classification Metrics in the yardstick package assume that the true value is a factor and the model returns a factor variable.
This is different behavior than for most explanations in DALEX, because when explaining predictions we typically operate on class membership probabilities. If we want to use Classification Metrics we need to provide a predict function that returns classes instead of probabilities.
So let’s prepare an explainer that has factor as y and the predict_function returns classes.
explainer_glm <- DALEX::explain(titanic_glm,
data = titanic_imputed[,-8],
y = factor(titanic_imputed$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#> variable mean_dropout_loss label
#> 1 _full_model_ 0.7977 lm
#> 2 gender 0.6334 lm
#> 3 class 0.7399 lm
#> 4 age 0.7875 lm
#> 5 sibsp 0.7923 lm
#> 6 embarked 0.7960 lm
#> 7 parch 0.7976 lm
#> 8 fare 0.7986 lm
#> 9 _baseline_ 0.5867 lm
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#> variable mean_dropout_loss label
#> 1 _full_model_ 0.7397031 lm
#> 2 gender 0.5619733 lm
#> 3 class 0.6788103 lm
#> 4 age 0.7268739 lm
#> 5 embarked 0.7346292 lm
#> 6 sibsp 0.7377564 lm
#> 7 parch 0.7393319 lm
#> 8 fare 0.7396152 lm
#> 9 _baseline_ 0.5051107 lm
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#> variable mean_dropout_loss label
#> 1 _full_model_ 0.1980 lm
#> 2 fare 0.1974 lm
#> 3 parch 0.1984 lm
#> 4 embarked 0.1986 lm
#> 5 sibsp 0.2035 lm
#> 6 age 0.2152 lm
#> 7 class 0.2525 lm
#> 8 gender 0.3658 lm
#> 9 _baseline_ 0.4144 lm
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.
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")#> Preparation of a new explainer is initiated
#> -> model label : Ranger Apartments
#> -> data : 1000 rows 5 cols
#> -> target variable : 1000 values
#> -> predict function : yhat.ranger will be used ( default )
#> -> predicted values : No value for predict function target column. ( default )
#> -> model_info : package ranger , ver. 0.14.1 , task regression ( default )
#> -> predicted values : numerical, min = 1899.414 , mean = 3488.861 , max = 6046.561
#> -> residual function : difference between y and yhat ( default )
#> -> residuals : numerical, min = -445.4107 , mean = -1.842099 , max = 736.702
#> A new explainer has been created!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#> variable mean_dropout_loss label
#> 1 _full_model_ 155.6184 Ranger Apartments
#> 2 no.rooms 316.5036 Ranger Apartments
#> 3 construction.year 392.5758 Ranger Apartments
#> 4 floor 431.6225 Ranger Apartments
#> 5 surface 524.9066 Ranger Apartments
#> 6 district 763.5492 Ranger Apartments
#> 7 _baseline_ 1206.6411 Ranger Apartments
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#> variable mean_dropout_loss label
#> 1 _full_model_ 0.983377709 Ranger Apartments
#> 2 district 0.326764600 Ranger Apartments
#> 3 surface 0.667962198 Ranger Apartments
#> 4 floor 0.775455846 Ranger Apartments
#> 5 construction.year 0.816108921 Ranger Apartments
#> 6 no.rooms 0.920990517 Ranger Apartments
#> 7 _baseline_ 0.001389768 Ranger Apartments
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.
Session info
#> R version 4.2.1 (2022-06-23)
#> Platform: x86_64-apple-darwin17.0 (64-bit)
#> Running under: macOS Big Sur ... 10.16
#>
#> Matrix products: default
#> BLAS: /Library/Frameworks/R.framework/Versions/4.2/Resources/lib/libRblas.0.dylib
#> LAPACK: /Library/Frameworks/R.framework/Versions/4.2/Resources/lib/libRlapack.dylib
#>
#> locale:
#> [1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
#>
#> attached base packages:
#> [1] stats graphics grDevices utils datasets methods base
#>
#> other attached packages:
#> [1] ranger_0.14.1 yardstick_1.0.0 DALEX_2.4.2
#>
#> loaded via a namespace (and not attached):
#> [1] tidyselect_1.1.2 xfun_0.32 bslib_0.4.0 purrr_0.3.4
#> [5] lattice_0.20-45 colorspace_2.0-3 vctrs_0.4.1 generics_0.1.3
#> [9] htmltools_0.5.3 yaml_2.3.5 utf8_1.2.2 rlang_1.0.5
#> [13] pkgdown_2.0.6 jquerylib_0.1.4 pillar_1.8.1 glue_1.6.2
#> [17] lifecycle_1.0.1 stringr_1.4.1 munsell_0.5.0 gtable_0.3.1
#> [21] ragg_1.2.2 memoise_2.0.1 evaluate_0.16 labeling_0.4.2
#> [25] knitr_1.40 fastmap_1.1.0 fansi_1.0.3 highr_0.9
#> [29] Rcpp_1.0.9 scales_1.2.1 cachem_1.0.6 desc_1.4.1
#> [33] jsonlite_1.8.0 ingredients_2.2.0 farver_2.1.1 systemfonts_1.0.4
#> [37] fs_1.5.2 textshaping_0.3.6 ggplot2_3.3.6 digest_0.6.29
#> [41] stringi_1.7.8 dplyr_1.0.10 grid_4.2.1 rprojroot_2.0.3
#> [45] cli_3.3.0 tools_4.2.1 magrittr_2.0.3 sass_0.4.2
#> [49] tibble_3.1.8 crayon_1.5.1 pkgconfig_2.0.3 Matrix_1.4-1
#> [53] rmarkdown_2.16 R6_2.5.1 compiler_4.2.1