What exactly is 'Decision Tree (Multiway)? What is its advantage?

Curious

If you can explain in simple words (beyond what is in the glossary). Thank you!

SGolbert

It generates decision tree models where the numeric attributes are split in fewer nodes but with multiple splits. In the normal decision tree numeric attributes are always split 2-way but appear in several nodes.

It's supposed to gain in interpretability, but I have tested it with the titanic data and it doen't seem to change much. Unless there are previous works that have used the multi-way approach, there is no reason to use it. For prediction Random Forest is the way to go, for interpretability an extra decision tree model can be used, or the Explain Predictions operator.

Regards,

Sebastian

varunm1

Hi @Curious

As @SGolbert explained they will have multiple splits for numerical attributes, but the advantages are minimal. One major advantage is the rate of information retrieval (fast) and updating the tree.

Thanks
Varun

Telcontar120

It also leads to shorter trees, all else being equal, since multiple splits can occur at the same node if the attribute it numerical, rather than requiring every distinct split to be only 2-way.  This can be a consideration in some implementations or with respect to a growth limit on the overall depth of the tree.

SGolbert

But in the end, the normal tree can be optimized regarding spread and deep. Also, I don't trust the implementation completely, it requires a decision tree at a subprocess, very strange. Better to stick to the normal, more tested one.