Inputs look unlike the training data, but no labels are available yet. What can you monitor?
Why B is the answer
Data drift
Drift monitoring compares inputs to a baseline and needs no labels.
In detail
- Data drift compares inputs with training data.
- It needs no labels.
- Add accuracy monitoring once labels arrive.
Use case
An illustrative scenario, not a real incident.
Loan outcomes take months to appear, so accuracy is unknown. Data drift monitoring compares inputs with training data now.
See it in different ways
flowchart TD A["Model monitoring"] --> B["Data drift: inputs changed"] A --> C["Prediction drift: outputs changed"] A --> D["Data quality: bad values"] A --> E["Performance: accuracy fell"]
stateDiagram-v2 [*] --> Trained Trained --> Registered: version and lineage Registered --> Deployed Deployed --> Monitored Monitored --> Retrained: drift alert Retrained --> Registered Deployed --> Archived: replaced
flowchart TD
Q{"What can you monitor?"}
Q -->|"No labels yet"| A0["Data drift"]
Q -->|"Labels available"| A1["Accuracy as well"]Key point: Data drift needs no labels.
Why the other options are wrong
- A. The time taken for the last training run
- Training time says nothing about inputs.
- C. Final accuracy against ground truth
- Accuracy needs labels, which are not available yet.
- D. The number of notebooks in the workspace
- Notebook counts say nothing about the model.