Oversee Model Governance PMI-CPMAI Practice Questions (Page 6)

Questions 26–30

1. 26expert · hard

   A telecom company's customer-churn model is retrained quarterly. The model's performance is stable, but the marketing team wants to launch a new campaign that targets a different customer segment. The model was trained on the general customer base. The team needs to decide whether to retrain before the campaign. What is the best approach?

   • ADelay the campaign until the next quarterly retraining cycle to avoid any disruption.
   • BKeep the current model and run the campaign, since the model's overall performance is stable.
   • CRetrain the model immediately on the new segment's data to ensure it is optimized for the campaign.
   • DEvaluate the current model's performance on the new segment's data, and retrain only if performance is below a defined threshold.
   Reveal Solution Discussion Select an answer first
2. 27expert · hard

   A credit-card issuer's fraud model is monitored with a dashboard that tracks AUC, precision, and recall. The dashboard shows that precision has dropped from 0.85 to 0.70 over the last month, but recall and AUC are stable. The input feature distributions are unchanged. The fraud team is concerned about the increase in false positives. What is the most likely cause and the best next step?

   • AThe model's decision threshold is no longer optimal for the current fraud-to-non-fraud ratio; re-evaluate the threshold using recent data.
   • BThe model's features have drifted, causing it to misclassify legitimate transactions; retrain with new features.
   • CThe model is suffering from model drift; retrain it immediately with the latest fraud data.
   • DThe monitoring dashboard is miscalibrated; ignore the precision drop and continue monitoring.
   Reveal Solution Discussion Select an answer first
4. 28application · medium

   A healthcare organization deploys a patient-readmission prediction model. The model's accuracy is stable, but the average prediction latency has increased from 200 ms to 900 ms over the past week, and the clinical dashboard times out. The team needs to identify the cause. Which monitoring approach is most directly useful?

   • ACompare the model's accuracy on a holdout set to confirm the model is still valid.
   • BTrack data drift on the input features to see if the patient population has changed.
   • CMonitor infrastructure metrics such as CPU, memory, and database query time alongside the model's latency dashboard.
   • DReview the model's feature importance values to see if the model is relying on different signals.
   Reveal Solution Discussion Select an answer first
5. 29foundation · easy

   Which of the following is an example of a key performance indicator (KPI) that would be tracked in model performance monitoring?

   • AModel inference latency
   • BThe number of data scientists on the team
   • CThe cost of cloud storage for training data
   • DThe version of the Python programming language used
   Reveal Solution Discussion Select an answer first
6. 30expert · hard

   A global bank deploys a loan-approval model. The model uses demographic features, and an internal audit reveals that the model's approval rate for a protected group is significantly lower than for other groups. The bank must comply with fair-lending regulations. The model's overall accuracy is high. What is the most appropriate governance action?

   • ARetrain the model with more data to improve its overall accuracy, assuming the disparity will resolve itself.
   • BConduct a fairness analysis to identify the source of the disparity, then adjust the model or its threshold to mitigate the bias while monitoring for unintended consequences.
   • CRemove all demographic features from the model to ensure fairness.
   • DDocument the disparity and continue using the model since its overall accuracy is high.
   Reveal Solution Discussion Select an answer first