in the decision tree discretiser, replace the attribute encoding pipeline by encoding dictionary

[solegalli]

Tree encoder: Change logic so that instead of encoder pipeline it has an encoding dictionary. I could return the output of the tree over the unique values to create a dictionary.

[Morgan-Sell]

Hi @solegalli,

I'm happy to work on this task, but what's the rationale?

A few more questions:

• How are you envisioning the encoding dictionary? What are the keys and values?
• How would a dictionary perform the fit()?

I think I'm missing something.

[solegalli]

This is another one that I come and go.

The DecisionTreeDiscretiser takes a continuous variable and replaces it by the outputs of a decision tree, trained on a single feature. Like this, it creates a monotonic relationship (at least in theory) between the encoded feature and the target, and also, returns a numeric value that can be used to train sklearn models.

That is the current functionality.

Now, originally, decision trees were used for discretization because they naturally find the best cutting points. And in discretization, we want cut points or otherwise interval limits, not predictions. The downside of returning intervals instead of predictions, is that we would still need to encode those predictions.

So I guess, we could expand this discretizer in various ways:

1. allow the discretizer to output predictions (current functionality, nothing to do)
2. allow user to round the values of the predictions (we would add a parameter called precision with the number of decimals after the comman, see other discretizers in feature engine)
3. expand the discretizer to return interval limits instead of predictions. And if we do this, then, like in our other discretizers, we would have the option to return the limits or to replace the intervals by integers.

So for 1 and 2, there is not much else to do, just introduce the param precision and round the predictions.

3 needs a bit of work, because we need to navigate the trees to find the limits. We've got the logic in this stackoverflow thread: https://stackoverflow.com/questions/75663472/how-to-obtain-the-interval-limits-from-a-decision-tree-with-scikit-learn

Hope that makes sense :)

Happy for you to take it.