問題文
A column of counts has 6 percent absent entries and the downstream model is a gradient-boosted tree. Which handling is defensible, and on what grounds?
選択肢
- Replace the absent entries with zero, because a count of zero is the natural reading of an absent count and using a single fixed value keeps the column's type unchanged while avoiding the extra column that a separate absence flag would introduce into the feature set.
- Leave the entries absent and add a boolean column recording that they were absent, so the model can use the fact of absence as information if it carries signal.
- Drop every row that has an absent entry in this column, since 6 percent is a small enough fraction that the loss is negligible, and a complete table keeps the training code simple and the feature list unchanged.
- Replace the absent entries with the column mean.