Question

You are preparing a churn classification dataset in which only about 8% of customers churned. You plan to perform a single 80/20 train/test split before training a model. A colleague warns that a plain random split could produce a test set whose churn rate differs significantly from the overall population, harming evaluation reliability. Which approach best addresses this concern while keeping the split reproducible?

Accepted Answer

Use a stratified split on the target label with a fixed random seed so both partitions preserve the ~8% churn proportion.

Answer

Oversample the minority class across the entire dataset before splitting so the classes become balanced.

Answer

Increase the test set to 50% of the data so the minority class is more likely to appear proportionally.

Answer

Sort the data by the target column before splitting so churned and non-churned rows are evenly distributed.