Pyramid Consulting uses a binary classification model in its candidate-job matching workflow to predict whether a submitted candidate will be accepted for client interview. The model is deployed in an internal Pyramid Consulting recruiter surface and assigns a score to each candidate profile so recruiters can prioritize outreach.
The team is debating which metric should drive model selection because different stakeholders care about different errors. Recruiters want fewer low-quality submissions, while account managers want to avoid missing candidates who would likely convert.
| Metric | Model A @ 0.70 threshold | Model B @ 0.40 threshold | Baseline Rules Engine |
|---|---|---|---|
| Precision | 0.91 | 0.63 | 0.58 |
| Recall | 0.42 | 0.81 | 0.54 |
| F1 Score | 0.57 | 0.71 | 0.56 |
| ROC-AUC | 0.79 | 0.86 | 0.68 |
| Positive rate predicted | 9% | 28% | 22% |
| Actual positive rate | 24% | 24% | 24% |
Client interview slots are limited, and each false positive wastes recruiter time and may hurt client trust. However, each false negative means Pyramid Consulting may miss a qualified candidate and lose placement revenue. You need to determine which metric should be prioritized for this use case and whether one model is clearly better.