30. A model flags suspicious transactions, but reviewers cannot examine them all. How would you choose a threshold?
MediumMachine learning basics
Technical
Show model answer
Clarify review capacity and the consequences of missed cases and unnecessary reviews. Evaluate precision, recall and expected workload across thresholds on representative held-out data, considering whether investigation delays change the value of an alert.
Choose an operating point with the accountable team and monitor changing volumes, case mix and outcomes. If labels arrive only for reviewed cases, explain that evaluation can be biased and design a suitable validation sample rather than assuming unreviewed transactions are negative.
What interviewers listen for
- Capacity and error costs
- Threshold tradeoff
- Selective-label bias