hard20 pts · first answer countsfeature-engineeringdata-leakage

Target encoding without leakage

You replace a high-cardinality "merchant_id" with the average fraud rate per merchant, computed on the full training set. Training AUC jumps far above validation AUC. What is the fix?

A.Compute the encoding out of fold, with smoothing towards the global mean for rare merchantsB.Use the average fraud rate from the test set insteadC.One-hot encode the merchant averagesD.Increase the model depth so it uses the feature better

Discussion 0

Join free to post in the discussion.

Target encoding without leakage

Discussion 0

Related questions