Practice — LightGBM (5 questions)
Advanced
Open
Pro
Diagnosing Overfitting and Choosing Regularization Parameters
Unlock this question →
Advanced
Open
Free
One New Feature Took CV AUC From 0.81 to 0.97 Permalink →
You add merchant_id (50,000 unique values) to a LightGBM fraud
model using mean target encoding: for each merchant, replace the ID
with that merchant's historical fraud rate, computed once over the
entire training set before doing anything else. You then evaluate
with standard 5-fold cross-validation. CV AUC jumps from 0.81 to
0.97. In production, AUC drops back to roughly 0.82.
What actually happened?
Share this question