Classregtree - implementing an additional constraint

Hej everybody,

I am using classregtree on a dataset with app. 200 entries from 25 data sources. I have been thinking of implementing an additional constraint into the tree growing process which demands that a new branch/leaf can only be grown if it contains samples from at least 3 different data sources. Is there an easy way to do this?

Thank you for your help,

John

Answer 1

[Ilya]

If you work in release 9b or later, the function searching for the optimal split in a tree node is coded in C for speed. The source code is not published. You could easily force every branch node in a tree to contain at least 3 classes; this part is coded in MATLAB. But you cannot impose the same requirement on a leaf node without modifying the C code or substituting it with your own (which I wouldn't recommend).

I don't know why you prefer a decision tree for your data and why you need at least 3 classes per node. If this is an attempt to improve the accuracy of predictions, there are better ways of accomplishing this.

[John Koestel]

Hej Ilya,

Thank you for the quick answer. However, it looks as if it was rather not so easy to implement it (I would also need it for leaf-nodes).

The tree is rather needed to have a tool to make rough predictions which can be intuitively understood. However, I am thinking of ways to decrease the probability that the tree is grown on proxies for the experimental conditions used in the various data sources.

thanks and cheers,

John