Control Decision Tree Depth

MATLAB's function,

fitctree

has name-value arguement to control the maximum number of branch node splits, the minimum leaf size and the minimum parent node size.

I would like to compare two different feature matricies with the same dataset. I want to evaluate the training error and CV error for each model built using the two different feature matrices. However, for a fair compaison, I would like to compute these errors with models of the same tree depth and I can't find a way to specify the number of levels I want the trees to have, nor can I find a pruning method that prunes by tree levels and not by nodes.

Does anyone have any ideas of what I can do? Thanks.

Answer 1

[Pratyush Roy]

There is no direct way to set the depth to which we want to grow the tree. This issue has been raised to the concerned people and they might be considered in the future releases of the MATLAB.

EDIT: I have received communication regarding a workaround. For tall arrays, one can use the 'MaxDepth' name-value property for setting the maximum depth to which we want to grow the tree.

The following link might be helpful:

https://www.mathworks.com/help/stats/fitctree.html#namevaluepairarguments

Hope this helps!

[Ryan Jones]

Thank you, that is good to know, so I don't keep trying to find methods of fixing the depth.

However, using the method using tall array results in a compact tree and you can't do basic operations such as cross-validation. Plus, I belive it takes longer to build than normal classification trees as it has to evaluate the tall arrays.