Surv => Regr methods have bugs?

[bblodfon]

Bugs

• Test row_ids are not correct
• During predict not the same features as in training (mlr3 new check triggered this, tests now fail):

Error: Learner 'regr.featureless' has received tasks with different columns in train and predict.

Main Pipeline

• https://mlr3proba.mlr-org.com/reference/mlr_graphs_survtoregr.html

Subsequent Pipeops

• https://mlr3proba.mlr-org.com/reference/mlr_pipeops_trafotask_regrsurv.html
• https://mlr3proba.mlr-org.com/reference/mlr_pipeops_trafotask_survregr.html
• https://mlr3proba.mlr-org.com/reference/mlr_pipeops_trafopred_regrsurv.html
• https://mlr3proba.mlr-org.com/reference/mlr_pipeops_trafopred_survregr.html

Building Pipeops for the above Pipeops

• https://mlr3proba.mlr-org.com/reference/PipeOpTransformer.html
• https://mlr3proba.mlr-org.com/reference/PipeOpTaskTransformer.html
• https://mlr3proba.mlr-org.com/reference/PipeOpPredTransformer.html

Action points

• Which methods do we need to keep?
• Update/simplify code for the above

[bblodfon]

After a session with Andreas, we concluded:

• Methods omit and delete for handling censoring are easy to manually implement
• mrl (mean residual lifetime) doesn't work currently (calls $survivalAntiDeriv from distr6), there is a good reference for it - chapter 2.4, Klein (2004) - no reduction technique is documented there though
• reorder seemed ad-hoc, will not work with newer mlr3 versions as we want to have the same features during training and prediction
• bj seems possible as a reduction technique (imputes censoring times), works, added a separate issue (Surv => Regr pipeline using Buckley-James imputation #420) for re-implementing it