Connect with `mlr3` [Suggestion]

[bblodfon]

Hi @ielbadisy ,

Great work with this package, it's good to finally have some DNN survival models in R!

We used to have Python-based wrappers via the survivalmodels R package in mlr3, but this package is no loner maintained, and many people have been asking for such models, see recent issue. So an R-based alternative is just what is needed!

Would you consider to spend some time trying to make an mlr3-compatible learner and connect to all things that make mlr3 nice? (general tuning via mlr3tuning, validation and internal tuning, mlr3proba, etc.). We have a lot of example survival learners, see the mlr3extralearners package. We welcome PRs and open-source collaboration!

BR, John.

[ielbadisy]

Hi @bblodfon,

thanks for the message! I really appreciate the interest from the mlr3 community. I'd definitely be interested in working towards making survdnn compatible with mlr3.

Before starting, would you be able to point me to a reference implementation or documentation about conventions followed in mlr3extralearners?

In particular, I'd appreciate some guidance on the recommended conventions, for example:

• whether implementing a standard LearnerSurv with the usual methods is sufficient for smooth integration with the ecosystem?
• how tuning is typically handled? whether learners should expose all hyperparameters directly to mlr3tuning, or if internal tuning utilities are ever wrapped?

Once I have a clear reference to follow, I'm happy to start with a minimal implementation and iterate from there.

Imad

[bblodfon]

Hi Imad, that's very nice to hear!

About conventions, we have a style wiki, but it hasn't been updated for some time. I would start by looking at some survival learners we have wrapped, see aorsf and xgboost reference implementations for two decent examples.

The extending doc is also useful to take a look at.

For the survival-related stuff, a good starter is the mlr3proba useR! tutorial, see here.

About tuning this book chapter should be a good start to get an idea of what can be done. In general we let mlr3tuning handle all the tuning, unless internally optimized, learner-specific methods exist (eg tuning of lambda in cv.glmnet is considerably faster and better to just use that than glmnet + mlr3tuning of lambda, another example is CoxBoost's internal tuning of the optimal learning rate and boosting rounds if I recall correctly). And if your DNN model can also do early stopping on a validation set, see this chapter on how it would look doing it via mlr3!

From a brief look at your code, I would just start with wrapping the survdnn() function on the .train() in a LearnerSurvDNN and the survdnn:::predict.survdnn on the .predict() following the above examples and see what is possible.

[ielbadisy]

John,

This is very helpful. I'll start working on this over the weekend and report back once I have a minimal prototype.

[ielbadisy]

Hi John,

I’ve opened mlr-org/mlr3extralearners#533 adding a minimal LearnerSurvSurvDNN.

Happy to iterate based on feedback.