For neural networks we have the universal approximation theorem which states that neural networks can approximate any continuous function on a compact subset of $R^n$.

Is there a similar result for gradient boosted trees? It seems reasonable since you can keep adding more branches, but I cannot find any formal discussion of the subject.

EDIT: My question seems very similar to Can regression trees predict continuously?, though maybe not asking exactly the same thing. But see that question for relevant discussion.

  • 1
    $\begingroup$ Good question! I could not find anything on that, but here are PAC bounds on decision trees. Try asking again on cstheory. $\endgroup$ – Emre Jun 7 '18 at 17:25
  • 1
    $\begingroup$ See here: projecteuclid.org/download/pdf_1/euclid.aos/1013203451. It is an old read. I believe it has what you are looking for. As far as I understand, in principle, they can. Let me know what you think of it. $\endgroup$ – TwinPenguins Jun 7 '18 at 18:27
  • $\begingroup$ I don't have a rigourous answer to your question but here are my thoughts: Trees are step functions and those can approximate any continuous function in $R^n$ $\endgroup$ – Smile Oct 19 at 17:04

Yes - create a region for each data point (i.e., memorize the training data).

Thus it is possible for gradient boosted trees fit any training data, but it would have limited generalization to new data.

| improve this answer | |

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.