In order to ensemble a decision tree, let me explain the specific situation. I have split the dataset in 5 sections, per protocol e.g. TCP, HTTP etc. Now I've trained a decision tree for each one and run this against my test dataset.

How do I go about combining the 5 prediction models. Do I....

a) combine the predicted output e.g. 1,0;1,1 for each section done separate and then run it against the test dataset to identify confusion matrix. This is actual class, predicted class. or
b) do I take the tree build and then add the additional ones to the tree model, in effect combining them.

Which method should be suitable, would option a even be a good solution?

  • $\begingroup$ Is it correct that you have just one target you are trying to predict? If so, why have you chosen to split the data by protocol? $\endgroup$
    – Tchotchke
    Jan 25, 2017 at 16:20
  • $\begingroup$ Thanks, I've ended up with my own method that uses a custom made threshold formula similar to bagging the predictions. Research paper will be published later this year. $\endgroup$ May 15, 2017 at 9:20

1 Answer 1


Popular ways to combine trained models are schema's based on voting: predictions are treated as weighted votes, the class with the majority of votes is selected as ultimate prediction, and/or stacking: predictions are treated as features in a newly trained model.

Depending on the domain at hand one of this approaches might help you. Stacking (depending on the metalearner of choice) introduces more freedom and might introduce a complexity that is not warranted by the concept you are searching for, while voting won't help much when identifying subdomain fitted for perticular problems in your search space.

Also when combining multiple classifiers it might be worth your while to try out some fundamentally different approaches.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.