Developing End-To-End ML Pipelines PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 3)

Questions 11–15

1. 11application · medium

   A team is building a model to predict housing prices. During training, they apply a log transformation to the target variable and standardize the features. At serving time, they plan to apply the same transformations using a saved preprocessing function. What is the primary benefit of this approach?

   • AIt increases the model's accuracy by making the data more normally distributed.
   • BIt reduces the model's training time because the data is already transformed.
   • CIt eliminates the need for data validation in the pipeline.
   • DIt ensures consistent data preprocessing between training and serving, reducing training-serving skew.
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A retail company trains a demand-forecasting model using a Vertex AI Pipeline that reads raw sales data from BigQuery, performs feature engineering (including a custom log transform and one-hot encoding), and then trains a model. The model is deployed to Vertex AI Endpoints for online predictions. During the first week of production, the online predictions are noticeably worse than the offline evaluation. The data science team suspects training-serving skew. Which action should they take to diagnose and prevent this issue?

   • AIncrease the number of epochs during training so the model better fits the training data and generalizes to the online traffic.
   • BRetrain the model using a larger holdout set to get a more accurate offline evaluation metric before deploying again.
   • CPackage the feature-engineering code as a reusable component and use the same component in both the training pipeline and the prediction container's preprocessing step.
   • DSwitch the online endpoint to use a smaller model instance type to reduce latency, which will improve prediction accuracy.
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   A fintech company is deploying a credit-risk model. During training, they normalize features using the mean and standard deviation computed from the training set. At serving time, they plan to normalize using the mean and standard deviation of the incoming request batch. What is the primary risk with this approach?

   • AThe model will experience training-serving skew because the normalization statistics differ between training and serving.
   • BThe model will be slower at serving time because it needs to compute statistics for each batch.
   • CThe model will overfit the training data because the normalization is not applied consistently.
   • DThe model will have a higher bias because the normalization is not applied to the training data.
   Reveal Solution Discussion Select an answer first
5. 14application · medium

   A data science team is developing a binary classification model to predict customer churn. They have a dataset with a highly imbalanced target (5% churn). They want to evaluate the model's performance on a holdout set that was never used during training or hyperparameter tuning. Which evaluation approach should they use?

   • AEvaluate the model on the training set to ensure it has learned the patterns well.
   • BCompute accuracy on the holdout set, since it is the most straightforward and widely understood metric.
   • CCompute precision, recall, and F1-score on the holdout set, and also plot the ROC curve and calculate the AUC.
   • DUse cross-validation on the entire dataset and report the average accuracy across all folds.
   Reveal Solution Discussion Select an answer first
6. 15application · medium

   A manufacturing company is building a predictive-maintenance ML pipeline for the first time. They have a small data science team and need to deliver a working pipeline quickly. They want to follow industry best practices for data validation, training, and evaluation without building everything from scratch. Which approach should they take?

   • AWrite a single monolithic Python script that runs all pipeline steps sequentially and schedule it with Cloud Scheduler.
   • BBuild a custom pipeline from scratch using Kubeflow Pipelines SDK to have full control over every component.
   • CUse Cloud Composer to orchestrate a series of Cloud Functions for each pipeline step.
   • DStart with a pre-built pipeline template from the Vertex AI Pipelines template gallery and customize it for their data and model.
   Reveal Solution Discussion Select an answer first