xgboost is the most famous R package for gradient boosting and it is since long time on the market.
In one of my publications, I created a framework for providing defaults (and tunability measures)
and one of the packages that I used there was xgboost. The results provided a default with
the parameter nrounds=4168, which leads to long runtimes.
Hence, I wanted to use the data used in the paper to set nrounds to 500 and optimize the other
parameters to get optimal defaults.
I used the same method as mentioned in the paper with the 38 classification datasets and obtained the following (rounded) defaults:
Performance on 29 regression datasets
Now I wanted to know how well this defaults perform on other tasks. So I took a collection of 29 regression tasks of different
domains, that I collected from different sources such as OpenML and compared my
defaults against the defaults of xgboost with nrounds=500. Code can be seen here.
The results were not really mindblowing. Here you can see performance graphs for two different measures (R-Squared and Spearman’s Rho)
using 10 times repeated 5-fold CV:
It is visible that my defaults perform better than the package defaults in most of the cases regarding the R$^2$ and also regarding
Spearman’s Rho (a nonparametric measure).
The average values can be seen in the following table:
The percentage of datasets where my new defaults are better than the old defaults is 66% for R$^2$ and 75% for Spearman’s Rho.
So all in all my new defaults do not provide a very big performance gain (on average 0.03 or R$^2$), but seem to be better
than the package defaults.
In my paper I could get an average improvement of 0.029 in AUC
for the classification datasets with my defaults. Also not a very big difference, so this is not very surprising.
The runtime was a bit longer with my defaults, but not really much longer. You can see the runtime below:
The longer runtime is mainly due to the max_depth parameter, which I set bigger (11 instead of 6).
All in all the datasets that I used were rather small, so bigger datasets would provide more interesting results.
In my next blog entry I will talk a bit about catboost and lightGBM.
Especially the first one promises to provide much better default performances than xgboost and I will compare it with the random forest
implementation ranger and the support vector machine implementation liquidSVM in their default state.