A support team predicts how long each phone call will last. Its boosted model starts by predicting 20 minutes for every call. The first XGBoost tree is then grown on the residuals of six past calls, where a residual is the actual duration minus the current prediction.
| Call | Read the help page first? | Actual minutes |
|---|---|---|
| A | no | 25 |
| B | no | 28 |
| C | no | 22 |
| D | yes | 21 |
| E | yes | 17 |
| F | yes | 24 |
XGBoost's regularization comes from two settings, here and .
A leaf's value is
A group of calls has a similarity score
and a split's gain is
The split is kept only if its gain is greater than 0. Otherwise the tree stays a single leaf holding all six calls.
The only candidate split is Read the help page first? After this tree, every call's prediction becomes its old prediction plus a learning rate times the value of the leaf it lands in.
What is call E's new predicted duration, in minutes? Round to 2 decimal places.