search

giuseppec / iml

iml: interpretable machine learning R package
https://giuseppec.github.io/iml/
Other
492 stars 87 forks source link

Support survival analysis tasks #172

Open RaphaelS1 opened 3 years ago

RaphaelS1 commented 3 years ago

Hi Christoph,

Thanks for making models more interpretable! Would you be happy to start supporting survival analysis in iml (via mlr3proba)? I'm happy to make a PR for this if you are.

Thanks!

christophM commented 3 years ago

What would you use as a prediction in the case of survival analysis? Do you have a quick proof of concept using a custom predict.fun for Predictor$new()?

RaphaelS1 commented 3 years ago

What would you use as a prediction in the case of survival analysis?

Agreed it isn't straight forward and it is messy but in general I'd say the only possible compatible prediction is a continuous relative risk ranking which you would then compute concordance on. Has little meaning in interpreting things like SHAPs (what does a relative increase of a relative prediction even mean) but is useful for feature importance.

Do you have a quick proof of concept using a custom predict.fun for Predictor$new()?

Can't Predictor just wrap mlr models?

Jack-FG commented 1 year ago

It would be nice to create ALE plots in addition to the pdp and marginal effects plots cabaple of being produced with these models: https://www.randomforestsrc.org/articles/partial.html

Random forest survival models