Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

After the awesome reception of my last blog about improvements for DALEX and DALEXtra, I couldn’t stand waiting for the next opportunity to share some details about new features with you. Sadly at some point, I realized that it’s necessary to implement them first. It took a while but finally after countless failed builds for GitHub actions, I’m glad to announce: “What's new in DALEX 2.1.0?”. With the new version, we will be able to choose which column of the output matrix will be taken into consideration while preparing explanations. Due to the character of that change, it obviously affects only classification tasks. But without further ado let’s move to working examples, and explaining what exactly that change means. For anyone interest in my previous blogs about DALEX here is the link.

### Penguins

Our journey starts in Antarctica. Quite a weird place for XAI stuff one may say. However, in today’s blog, 3 different species of penguins exactly from that continent will serve as our companions. penguins is a dataset coming from the palmerpenguins R package. As README states, the data were collected and published by Dr. Kristen Gorman and contain information about 344 penguins living in 3 different islands in the Plamer Archipelago. Authors of the package see penguins as an alternative for iris, let’s see!

library(palmerpenguins)
data_penguins <- na.omit(palmerpenguins::penguins)


### Predict function target column

DALEX 2.1.0 brought a new parameter to explain a function which is predict_function_target_column. It allows users to actually steer the flow of the model’s response that is taken into consideration in explaining classification task models. In previous versions, whenever the default predict_function was used, DALEX was returning the second column of the output probabilities matrix for binary classification, and the whole matrix of probabilities for multiclass classification. While that default behavior was preserved, now bypassing predict_function_target_column parameter we can force DALEX to take the column that we are interested in without being forced to pass the custom predict_function. What’s even more fantastic is that you can do it not only for binary classification models but also for multiclass classification changing the task to one vs others binary classification. It’s a super handy tool whenever one of the classes in your multiclass task is far more important than the others and such analysis, made from both perspectives may be useful. The usage of that parameter is super useful. It accepts both numeric and character inputs and should be understood as either order of the column in the probabilities matrix or the name of the column that should be extracted. To be even more precise, the parameter’s value is used directly to index the column. Keep in mind that some of the engines like gbm returns a single vector for multiclass classification. The change does not affect such models.

### Creation of model and explainer

That being said it’s high time to make some models and show new functionalities in practice. For that purpose, we will need two predictive models. Let them be simple, performance is not what we seek for today. The first model is going to be a binary classification. For this purpose, we will need to create a new variable, is_adelie which I think is really self-explanatory. The second model will be a multiclass classification for species variable. The engine behind both of them will be a ranger.

library(“ranger”)
library(“DALEX”)
model_multiclass <- ranger(species~., data = data_penguins,
probability = TRUE, num.trees = 100)
explain_multiclass_one_vs_others <- explain(
model_multiclass,
data_penguins,
data_penguins$species == “Adelie”, label = “Ranger penguins multiclass”, predict_function_target_column = “Adelie”) explain_multiclass <- explain( model_multiclass, data_penguins, data_penguins$species,
label = “Ranger penguins multiclass”,
colorize = FALSE)

model_binary <- ranger((species==”Adelie”)~., data = data_penguins,
probability = TRUE, num.trees = 100)
explain_binary <- explain(
model_binary,
data_penguins,
label = “Ranger penguins multiclass”,
colorize = FALSE,
predict_function_target_column = 2)

As you see the usage is very simple! Therefore let's explain some models!

Model performance is always a good place to start. An important note is that multiclass models with passed predict_function_target_column parameter are treated as standard binary classification, so measures for binary will be displayed. Let’s see the difference

(mp_one_vs_others <-
model_performance(explain_multiclass_one_vs_others))
(mp <-
model_performance(explain_multiclass))

Another way that lays in front of us with the new option is calculating feature importance using different measures. Default measures are loss in cross-entropy for multiclass and one minus AUC for binary. With the change we can in fact calculate which features are most important for predictions of one specific class, isn’t that amazing?

fi_one_vs_others <- model_parts(explain_multiclass_one_vs_others)
fi <- model_parts(explain_multiclass)
plot(fi_one_vs_others)
plot(fi)

I’m quite sure you are familiar with how Predict Profile and Model Profile explanations handle multiclass models. They simply calculate explanations for each level of the y and then combine them together on the plot. But we are not only interested in profiles or breakdowns for all of the levels, right? Here comes another field when a new parameter can be utilized, we can simply choose which parameter should be included.

pdp_one_vs_others <- model_profile(
explainer = explain_multiclass_one_vs_others,
variables = “bill_length_mm”)
pdp <- model_profile(
explainer = explain_multiclass,
variables = “bill_length_mm”)
plot(pdp_one_vs_others)
plot(pdp)

bd_one_vs_others <- predict_parts(
explainer = explain_multiclass_one_vs_others,
new_observation = data_penguins[1,])
bd <- predict_parts(
explainer = explain_multiclass,
new_observation = data_penguins[1,])
plot(bd_one_vs_others)
plot(bd)

### Summary

That will be enough for today! I hope you are as excited about a new feature as I am. I didn’t focus on methods used today, therefore if you want to know more about pdp, iBreakDown, or feature importance, I encourage you all to visit XAI tools page, where you can find an overview of different solutions for XAI in R and Python. There is also an excellent book Explanatory Model Analysis referring to that subject. As always, in case of any questions or problems feel free to open issues at https://github.com/ModelOriented/DALEX or https://github.com/ModelOriented/DALEXtra repos. We look for your suggestions regarding the future of our software.

If you are interested in other posts about explainable, fair, and responsible ML, follow #ResponsibleML on Medium.

In order to see more R related content visit https://www.r-bloggers.com/

DALEX 2.1.0 is live on GitHub! was originally published in ResponsibleML on Medium, where people are continuing the conversation by highlighting and responding to this story.