Deduplicate records before creating train and test: Which approach should replace the current evaluation?
Deduplicating before partitioning prevents shared records from contaminating both training and held-out evaluation.
The question
A customer-service team reports excellent test performance, but an audit finds duplicated customer conversations split between training and testing. The approval constraint is a credible estimate for new customer conversations. Which approach should replace the current evaluation?
Preparing for AIGP? Take the free 5-min readiness quiz →
- Deduplicate records before creating train and test partitions. ✓Removing duplicates before splitting prevents the same customer conversation from informing both training and evaluation, restoring credible testing.
- Add more held-out rows without reviewing duplicate relationships.Additional rows do not cure leakage if related conversations continue crossing training and testing partitions.
- Randomly reshuffle the existing rows before retraining.Random reshuffling can place related duplicates across partitions again, preserving leakage and overstating performance on new conversations.
- Keep duplicates but report separate metrics for short and long conversations.Length-based reporting does not remove duplicated information shared across partitions or establish performance on genuinely new conversations.
The trap
Inspect relationships among records before splitting; independence requires more than assigning individual rows randomly. How to remember it
Deduplicating before partitioning prevents shared records from contaminating both training and held-out evaluation.
How many of these would you get right?
One of 1581 AIGP questions on Certsqill. Take a free five-minute check and see your score per domain — not one number, but which section to open tonight.
Test your AIGP readiness — freeMore Understanding How to Govern AI Development questions
- Use a representative: Which approach satisfies that requirement? →
- Measure each subgroup's false-positive rate against 5%: Which evaluation approach best satisfies that →
- Use a cost-sensitive threshold: Which approach best addresses both the asymmetric error costs and limited →
- All 426 Understanding How to Govern AI Development questions →
Part of the Certsqill AIGP question bank · Understanding How to Govern AI Development ·
Every answer, right and wrong, comes with its own explanation.
