Simplify `TaggedTable`

[lars-reimann]

Is your feature request related to a problem?

We've already spent a lot of time overriding Table methods in TaggedTable just to keep the tagging. There are also more issues waiting, e.g. #644. Moreover, we also want to connect other types of input and output data, e.g. images to a table, so TaggedTable is just a special case of a more abstract Dataset concept.

Desired solution

• Don't inherit Table in TaggedTable.
• Get rid of all the overridden Table methods.
• Table.tagColumn can stay. It's now the final step before passing the data to an ML model.
• Regressor.fit and Classifier.fit could still return a TaggedTable or maybe just a Table?

Possible alternatives (optional)

No response

Screenshots (optional)

No response

Additional Context (optional)

No response

[lars-reimann]

There's also a strong correlation between TaggedTable and the InputConversionTable. Both expect names of features and the name of the target.

[Gerhardsa0]

There's also a strong correlation between TaggedTable and the InputConversionTable. Both expect names of features and the name of the target.

I aggree, also with TimeSeries and test:inout_conversion_time_series. Maybe we should let the NN detect the type and then cast, because a lot of information is stored there.
I will make a prototype in my PR #615

[lars-reimann]

🎉 This issue has been resolved in version 0.22.0 🎉

The release is available on:

• v0.22.0
• GitHub release

Your semantic-release bot 📦🚀