Microsoft AI-300: Implement machine learning model lifecycle and operations Practice Questions

Using a weighted or macro-averaged metric such as AUC_weighted or average_precision_score_weighted ensures that Automated ML's model selection is driven by performance on the minority class, rather than being dominated by the majority class as with raw accuracy. This is the recommended primary metric change for imbalanced classification. Setting featurization mode to 'auto' enables Automated ML's automatic featurization pipeline, which includes class imbalance detection and mitigation techniques such as oversampling (SMOTE) and class-weight adjustments. There is no enable_class_balancing parameter in the Azure Machine Learning Automated ML SDK v2 API; this is a distractor reflecting a plausible but non-existent setting. Setting n_cross_validations to 1 effectively disables cross-validation, which would produce unreliable performance estimates especially on an imbalanced dataset, and is counterproductive. LogisticRegression supports class_weight='balanced' natively and is one of the models that handles imbalance reasonably well; blocking it would reduce the model search space without benefit.

mlflow.autolog() enables automatic logging of parameters, metrics, and artifacts for supported frameworks without requiring explicit log calls in the training script. mlflow.log_param() requires manual calls for each parameter and does not log metrics or artifacts automatically. mlflow.set_tracking_uri() configures where runs are stored but does not enable automatic logging. mlflow.start_run() with nested=True is used for hierarchical run organization, not for automatic logging.

Bayesian sampling intelligently selects the next set of hyperparameters based on the results of previous runs, converging to good solutions much faster than random or grid search in large, expensive search spaces — ideal when runs are costly and time is limited. Bandit early termination policy cancels any run whose primary metric is not within a specified slack factor of the best run so far, immediately freeing compute for more promising configurations. — Grid search exhaustively tries every combination and is computationally infeasible with 20 hyperparameters and a 4-hour-per-run budget. — Median stopping terminates runs whose performance falls below the running median of completed runs; it is less aggressive than Bandit and less well-suited to very expensive runs with constrained budgets. — Setting max_total_runs to a very large number removes a key cost constraint and contradicts the requirement. — A sequential sweep job runs one trial at a time, eliminating parallelism and making the already expensive search even slower.

Calling mlflow.set_experiment() designates the target experiment, and mlflow.start_run() opens a new run context under that experiment, ensuring all subsequent mlflow.log_param(), mlflow.log_metric(), and mlflow.log_artifact() calls are scoped correctly and visible in the Azure Machine Learning studio experiment view. mlflow.log_param() alone logs a single parameter but does not create or scope a run context, so subsequent logs may end up in an unintended active run. mlflow.autolog() can automatically capture framework-specific metrics but still requires an active run context started via start_run() to be reliably scoped to the correct experiment; calling it alone at module level without an explicit run context is not the correct pattern for controlled experiment tracking. mlflow.register_model() is used to register a trained model in the model registry after training is complete, not to initialize experiment tracking at the start of a script.

Asynchronous parameter server training allows worker nodes to push gradient updates to a central parameter server and pull updated weights without waiting for all other workers to complete their gradient computation, eliminating the synchronization barrier and reducing idle time. — Synchronous data parallelism with NCCL all-reduce (e.g., standard PyTorch DDP) is efficient but requires all workers to synchronize gradients each step — the very bottleneck described. — Horovod ring-allreduce is also a synchronous method; every worker must participate in the ring before weights are updated. — Model parallelism splits the model graph across devices and is used when a model is too large to fit in a single GPU's memory; it does not address gradient synchronization bottlenecks from data-parallel training.