Best practice for "Imputers"

[preinh]

What about imputers?

Do you have any recommendation or best practice for address them?

• When they need to be applied for training data as well before to produce predictions.
• When they need to be distinct for categorical and continuous values.

[Philoso-Fish]

We handled imputers in our catalog and experiments for CARLA by removing these samples. Our datasets were, luckily big enough.
I think most of our recourse methods cannot handle missing data, such that imputers have to be dealt with before.

At the moment we don't have a best practice strategy to handle them, maybe I would start with techniques mentioned in https://medium.com/fintechexplained/what-are-imputers-in-data-science-b72f8308322b

[preinh]

Thanks @Philoso-Fish , I actually handled them particularly at the Data object. As well exposing the 'fitted' methods for later usage, such getting counterfactuals for later inferences. It worked here but I'm still struggling to generalize it as an skeleton or something like that.

Thanks also for the reference. I'll follow it and growth a little bit, unfortunately not as much as our current real dataset / business problem. :-)

Btw, congratulations for CARLA initiative of benchmarking recourse methods.