[SOLVED] Model Interpretability for Survival task in MLR3

Model Interpretability for Survival task in MLR3

I have performed previously model interpretation (pdp) on my survival models in MLR. However, I am unable to do it in MLR3 due to the "Predictor" object that is not accept models from type survival.

I have attached below a sample code from MLR, is there anyways to do model interpretation in MLR3?

mod = train(lrn,surv.task)

task.pred = predict(mod,newdata = traindata)

getLearnerModel(mod)

pd = generatePartialDependenceData(mod,surv.task,c("X","Y","Z"))

plotPartialDependence(pd)

as for c("X","Y","Z")) #These are three random features (I want to be able to do it for all features and for selected ones like in MLR)

Solution

to my knowledge mlr3 cannot directly apply interpretable machine learning methods to survival models. However, there is the survex package, which can perform a plethora of explainability analyses and is compatible with mlr3proba (survex package GitHub link)

Here is a minimal code example how you can generate PDPs from a fitted mlr3proba model, you can check the package vignettes for more information:

library(mlr3proba)
library(mlr3extralearners)
library(mlr3pipelines)
library(survex)
library(survival)

veteran_task <- as_task_surv(veteran,
                     time = "time",
                     event = "status",
                     type = "right")
ranger_learner <- lrn("surv.ranger")    
ranger_learner$train(veteran_task)
ranger_learner_explainer <- explain(ranger_learner, 
                     data = veteran[, -c(3,4)],
                     y = Surv(veteran$time, veteran$status),
                     label = "Ranger model")

model_profile <- model_profile(explainer)

plot <- plot(model_profile, 
             variables = "celltype")
plot