Question

You deployed a fraud-detection model to a managed online endpoint. Actual fraud labels (ground truth) only become available 30 to 45 days after a transaction is scored, once chargebacks are confirmed. Your team wants to monitor the model's classification accuracy and F1 score in production and alert when they degrade below a threshold. Which approach correctly configures Azure Machine Learning model monitoring to compute these performance metrics?

Accepted Answer

Configure a model performance monitoring signal that joins collected inference data with the delayed ground-truth dataset on a correlation ID, and schedule the monitor to run on a cadence aligned with when labels arrive.

Answer

Configure a data drift signal comparing production inputs to the training baseline, since drift in inputs is equivalent to a drop in accuracy and F1.

Answer

Enable a prediction drift signal on the model's output distribution, which reports accuracy and F1 automatically without needing ground-truth labels.

Answer

Configure a data quality signal to track null rates and type mismatches, which surfaces accuracy and F1 degradation as quality violations.