フリー問題

問題文

A pipeline imputes missing numeric entries with the column mean. Where must that mean be computed to avoid leakage?

選択肢

1. From the training subset only, with the same stored value then applied to the evaluation subset and to production records.
2. From each subset separately, using its own mean, so every subset is filled with a value that matches the distribution of that same subset.
3. From the production records at inference time, so the filled value tracks the traffic arriving at that moment and follows it as it drifts.
4. From the combined training and evaluation subsets, because using more rows gives a more accurate estimate of the column mean and a more accurate imputation can only help the model generalize better.