Context

SafeHire uses a binary classification model to flag potentially fraudulent job applications for manual review before employers can contact candidates. The current gradient boosting model performs well on recall, but operations is escalating that too many legitimate applicants are being incorrectly blocked, creating recruiter friction and support costs.

Current Performance

The Problem

In the last 30 days, the model reviewed 50,000 applications. It caught most fraudulent applications, but the business estimates that false positives are now the dominant source of operational cost because each incorrect flag triggers manual review, delays candidate response time, and causes some employers to abandon the platform.

Task

1. Interpret the current metrics and explain why this model is misaligned with business cost.
2. Use the metrics and confusion matrix implications to diagnose whether the model is too aggressive.
3. Recommend how you would optimize the model when false positives are much more expensive than false negatives.
4. Propose specific threshold, calibration, and evaluation changes.
5. Explain what additional error analysis you would run before retraining.

Constraints

• Manual review team can handle at most 2,200 applications per day.
• Product leadership does not want recall to fall below 0.75.
• Any proposed change must be measurable in a 2-week offline validation cycle.