Multi-target regression - one model or multiple submodels

[krzjoa]

Hi,

I have a question regarding the multi-target regression tasks. I know how to do that with TFDF (passing a list of MultiTaskItem(label=target_column, task=tfdf.keras.Task.REGRESSION) objects), but I'm curious, how it's implemented.

Does YDF create a separate submodel for each target variable or use one model to predict all the target variables simultaneously?

[rstz]

Hi,

This is a part of the library that's under active development, so the answer might change soon. The last release of TF-DF still has independent models for each objective. However, a recent commit to YDF allowed Multi-task learning where the output of models trained as secondary are used as input for the "primary" model. This (and possibly more) changes to Multi-task models will land withe next TF-DF version.