A SAS developer builds a gradient boosting model using PROC TREEBOOST with the NTREES=200, SHRINKAGE=0.05, and SUBSAMPLE=0.8 options. What does SUBSAMPLE=0.8 control in stochastic gradient boosting?
Select an answer to reveal the explanation.
Short Explanation
Here's the deal — SUBSAMPLE=0.8 tells SAS to randomly sample 80% of the training observations, without replacement, to fit each successive boosting tree. That row-level randomness is what makes it stochastic gradient boosting, and it helps reduce overfitting.
Full Explanation
The correct choice is the fraction of training observations randomly sampled (without replacement) to train each successive boosting tree. In stochastic gradient boosting, SUBSAMPLE= specifies the fraction of training rows randomly selected to fit each successive tree. With SUBSAMPLE=0.8, each tree is trained on 80% of training observations, introducing stochasticity that improves generalization and reduces overfitting. The variable-sampling choice describes column (feature) subsampling. The train/validation split choice is a data preparation step, not the SUBSAMPLE option. The gradient-computation subsampling choice is wrong because that parameter does not control gradient computation cost.