A hospital retrains the same kind of mortality-risk model every six months. One stored patient record receives an estimated risk of 18%, then 29%, then 21%.
The patient did not change. The training data did—and a decision tree can react sharply to small differences in that evidence.
What if, instead of trusting one tree, we let many varied trees answer and average their predictions?
In this week’s interactive article, you’ll keep one input fixed, watch predictions move across training samples, and test how bagging narrows that distribution. You’ll also see why adding more trees eventually helps less when their predictions move together.