フリー問題

問題文

A labeled table has 40 000 rows and a binary target that is positive in 1.8 percent of them. The model will be evaluated on a holdout. Which split type is appropriate?

選択肢

1. A stratified split that preserves the positive rate in both parts, so the holdout contains enough positives for the score to be stable.
2. A split by row order, taking the last 20 percent as the holdout.
3. No split; report the training score with cross-validation only.
4. A plain random split with a fixed seed, because fixing the seed makes the split reproducible and reproducibility is the property that matters for a fair comparison between candidate models, while the class balance will match the population in expectation for a table of this size.