I want to create a classification tree in Matlab using classregtree. However, a lot of data comes from wide histograms. I’ve noticed that when I add more histograms, the tree becomes worse. Is it possible to tell Matlab that it should group certain columns so it won’t threat all the values from the histogram separately, but evaluates them as one?
Edit: to clarify it a little bit more, I’ll provide an example from my current project. I have a dataset ‘A’ and a dataset ‘B’. B represents a histogram, and A represents all sorts of data. When I use A as the trainingset for the tree, and I use that set to test the tree, I get a score of 155/220 elements were correct. If I do the same for B, I get 97/220. However, if I add B to A, I get 145/220, which is less than 155/220. I think this is because classregtree doesn’t know that all the values from B actually represents a single variable.
It’s a long time ago I posted this question, but I have found the answer. It is not possible to group certain elements in a regression/classification tree, however there is a solution which provides similar result.
Matlab can also tell you how certain it is that an answer of a tree is correct using the
testandnodeerrfunctions. Using this information it is possible to combine multiple trees. So instead of putting all the data in one tree, you can create multiple trees. Now you can execute these trees independently, and obtain multiple, perhaps different, results. Using the error values of each result you can calculate which answer is most likely correct.