Question 30
Domain 4: Deploy and operationalize machine learning solutionsYou are using Azure ML SDK v2 and Fairlearn to evaluate and mitigate bias in your model, applying the Equal Opportunity Parity constraint. You employ the `ThresholdOptimizer` post‑processing algorithm. For each statement below, select “Yes” if true, otherwise “No”:
Correct answer: C
Explanation
No. `ThresholdOptimizer` is a post-processing mitigation method in Fairlearn, so it operates on model outputs rather than on an Azure ML inference cluster. Equal Opportunity Parity requires equal true positive rates across groups, and the algorithm learns group-specific thresholds to satisfy that constraint.
Why each option is right or wrong
A. Attached compute
B. Compute cluster
C. Inference cluster ✅
Fairlearn’s `ThresholdOptimizer` is a post-processing algorithm defined to fit on predictions/decision scores after training, not on Azure ML compute infrastructure; the relevant Fairlearn API is `ThresholdOptimizer(constraints='equalized_odds'/'true_positive_rate_parity', ...)`, which adjusts group-specific thresholds at inference time. Azure ML inference clusters are deployment targets for serving models, but they are not the mechanism that enforces Equal Opportunity Parity under this algorithm, so selecting the cluster as the mitigation object is incorrect.
D. Compute instance **Correct answer:** C. Inference cluster **Explanation:** An **Inference cluster** is the only Azure ML compute option designed to host real‑time web service endpoints under full Azure management. It provides built‑in autoscaling, high availability, and integration with container orchestration: - **Azure Kubernetes Service (AKS)**: ideal for production scenarios, offering robust autoscaling, rolling upgrades, and SLA‑backed uptime. - **Azure Container Instances (ACI)**: suited for development and testing or low‑volume endpoints, allowing quick, serverless deployment without cluster management overhead. Other compute targets do not meet the requirements: - **Compute clusters** (AML compute) are optimized for batch training and pipeline jobs, not for serving live HTTP requests. - **Compute instances** serve as interactive notebooks and development workspaces, lacking the orchestration needed for scalable inference. - **Attached compute** refers to external resources you bring into Azure ML, which you then manage yourself rather than letting Azure handle scaling and availability. By choosing an Inference cluster, you gain a fully managed environment that automatically provisions and scales resources, integrates with your CI/CD pipelines, and ensures low‑latency, reliable real‑time scoring. **References:** - [Deploy a real-time endpoint with Azure ML Designer](https://learn.microsoft.com/en-us/azure/machine-learning/tutorial-designer-automobile-price-deployment) - [Azure ML compute targets overview](https://learn.microsoft.com/en-us/azure/machine-learning/concept-compute-target) - [Configure inference clusters in Azure ML](https://learn.microsoft.com/en-us/azure/machine-learning/how-to-deploy-azure-kubernetes-service)