Feature: support SurvSHAP computation with {treeshap}

[kapsner]

This PR adds the feature to compute (global and local) SurvSHAP values with the {treeshap} algorithm, which drastically decreases computation time for ranger survival models as compared to using the currently implemented {kernelshap} algorithm.
Furthermore, this PR enables to compute SurvSHAP with models that were trained with data objects of type matrix.

Unit-tests have been added as well.

Addresses #75 .

The only thing I've just identified as a potential challenge is that {treeshap} is not on CRAN yet - are there plans to bring {treeshap} to CRAN in the near future?

[kapsner]

Unit-test showed error regarding missing output_type argument when using "exact_kernel" -> this is fixed with 0e8c85c

[kapsner]

Hi @mikolajsp and @krzyzinskim ,
have you already had some time to have a look at this PR? Pls. let me know if you have any further questions.
Best, Lorenz

[krzyzinskim]

Hi @kapsner,
it looks great! Thanks so much for your contributions.

The only thing I've just identified as a potential challenge is that {treeshap} is not on CRAN yet - are there plans to bring {treeshap} to CRAN in the near future?

As you mentioned treeshap is not on CRAN, but we have already started to address this (you can look at my PR). I have slightly improved your implementation and added you also as contributor to the treeshap package. Once the stable version is ready, we will also accept this pull request too. Hopefully, it will be this week.

[kapsner]

Hi @krzyzinskim ,

thanks a lot. The PR looks good and thanks for mentioning me as a contributor of the treeshap package!
Let me know if I can contribute anything else regarding the CRAN-submission of treeshap!

[krzyzinskim]

@kapsner, treeshap is already on CRAN, so that one issue is settled 🎉 Thanks for your willingness to help!

Unfortunately, I made one silly mistake while trying to generalize the functions for explanations also for CHFs. I've already managed to find the mistake and fixed it. However, using the version available on CRAN results in the explanations for the survival curves not being calculated properly. Sorry for that.

With CRAN's policy in mind, I don't want to do a new submission straight away. Maybe I will use this week or two to try to add support for another survival model (maybe randomForestSRC) and some other enhancements. Hope it's ok for you.

In the meantime, I'll merge this PR to a separate branch to work in parallel on the integration with survex.

[kapsner]

@krzyzinskim ,
Thank you very much for all of your efforts and congratulations that treeshap is on CRAN now 🎉

I understand the topic with the CRAN policy well - in the end, they want to avoid more than 6 or 7 CRAN submissions per year. I had that once with one of my packages, however, providing a reasonable explanation in the cran-comments allowed us to also update the package even more often (although this should be an exception).

Waiting for another survival model makes indeed sense 👍

For the treeshap package, I've also created another small PR adding my ORCID (ModelOriented/treeshap#32).

Best,
Lorenz