Repeated records from the same customer share information. A random row split can therefore evaluate a familiar person when deployment requires predicting a new one. Group validation keeps an entity entirely on one side of each split. You will inspect overlap, calculate grouped losses and count independent held-out entities. Group folds don't automatically balance every target distribution, so inspect event rates as well. The correct split depends on the deployment question: returning subscribers and first-time subscribers need different evidence. In this lesson, six fictional records belong to each subscriber and the target use case is a subscriber absent from training.