2
$\begingroup$

I found this document which compare some learning methods and I don't understand this table : enter image description here

Gradient boosting has a better intepratability score than CART. How is it possible ? I thought gradient boosting was an ensemble method, so we have to take a look at all the trees in order to understand how the classification is built.

Dou you think there is a mistake here ?

$\endgroup$
1
  • 1
    $\begingroup$ This display is tough on people who are red-green colour blind. A green dot is good and a red dot is poor. Poor design. $\endgroup$
    – Nick Cox
    Commented Dec 8, 2013 at 18:31

2 Answers 2

4
$\begingroup$

Im sure this is a typo. This document appears to be lecture notes from from Dr. Hastie from Stanford. Please look at Dr. hastie's book by following the link below at pg 351 table 10.1 has an accurate comparison and comprehensive background for machine learning methods, but does not have gradiant boosting method compared.

http://www.stanford.edu/~hastie/local.ftp/Springer/OLD/ESLII_print4.pdf

$\endgroup$
1
  • $\begingroup$ I would say that I definitely agree with the document. When trying to interpret RF have a forest of often hundreds of trees to look through. When using GBM you end up with a single we trained tree. Thus you need only look at one tree as opposed to the incomprehensible 100s. $\endgroup$
    – josh
    Commented Sep 17, 2016 at 22:30
0
$\begingroup$

(1) Interpretability is fairly subjective.
(2) Who authored the document? Do they have an objective view on the subject?
(3) Some effort has been made to make boosting more interpretable (partial dependence, relative importance) and easy as a result to overestimate interpretability. (One appreciates what you've put effort into achieving)
(4) What are the comparisons? No MARS or logistic regression?
(5) The effect/importance of a single variable may not always be apparent in CART (?).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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