3

Both the Bradley-Terry model and Gradient boosted trees can be used to learn a ranking from pairwise comparisons (e.g. with libraries choix and XGboost).

How do they relate to each other? Is there any situation in which one might be preferable to the other?

I would say that gradient boosted trees are more general than Bradley-Terry, so the only advantage of using Bradley-Terry might be in terms of training time / possibly robustness to overfitting, but I would like to hear other opinions on this.

vman
  • 173
  • 5
  • I guess one important difference is that XGboost is trained on features and Bradley-Terry no - although this R package for Bradley-Terry https://cran.r-project.org/web/packages/BradleyTerry2/ also can be trained on features (or explanatory variables, to use the package terminology) – vman Dec 02 '19 at 06:04
  • I have used both. Hands down the more important thing was out of sample evaluation. That goes for any model used. BT models are great, and GBM models are great and one does not always dominates that other. That said, for example, I have seen no Spark implementation of BT models in which case we are constrained by tools' availability. – usεr11852 Dec 18 '19 at 14:17

0 Answers0