Building Models Given the Task Considering Cost, Complexity, Latency, and Scalability PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 5)

Questions 21–25

1. 21foundation · easy

   A data scientist wants to explain why a complex model made a specific prediction for a single customer. Which technique is designed to provide local, per-instance explanations?

   • AHyperparameter tuning
   • BModel accuracy on a test set
   • CFeature importance from a tree model
   • DSHAP (SHapley Additive exPlanations)
   Reveal Solution Discussion Select an answer first
2. 22expert · hard

   A social media company wants to detect hate speech in user posts in real time. They have a large dataset of labeled text and need to balance accuracy, latency (under 200ms), and cost. They are considering a fine-tuned BERT model, a logistic regression on TF-IDF features, and a rule-based system. The rule-based system is too inaccurate. The logistic regression is fast but less accurate. The BERT model is accurate but may be too slow and costly. What is the most appropriate approach?

   • AUse a rule-based system with a large list of offensive words.
   • BUse the full BERT model on a CPU endpoint.
   • CUse logistic regression on TF-IDF because it is fastest and cheapest.
   • DUse a distilled BERT model deployed on a GPU endpoint.
   Reveal Solution Discussion Select an answer first
4. 23foundation · easy

   An ML model predicts equipment failure from sensor data. The predictions are needed daily to generate maintenance reports, and the team can tolerate a few hours of delay. Which deployment strategy is most appropriate?

   • AOnline prediction
   • BBatch prediction
   • CEdge deployment
   • DStreaming prediction
   Reveal Solution Discussion Select an answer first
5. 24expert · hard

   A global e-commerce company needs to build a product recommendation model. They have data in BigQuery across multiple regions and require that data not leave the region where it is stored. The team has data scientists who can write custom Python code, but they want to avoid managing training infrastructure. They also need to serve predictions with low latency. Which combination of products is most appropriate?

   • AUse BigQuery ML to train a model on aggregated data in one central region.
   • BUse Vertex AI AutoML Tables to train a model in one region and deploy globally.
   • CUse Vertex AI Pipelines with training jobs in each region and deploy online endpoints in each region.
   • DUse a pre-trained LLM hosted globally to generate recommendations.
   Reveal Solution Discussion Select an answer first
6. 25application · medium

   A media company wants to build a recommendation system for articles. They have a large dataset of user interactions in BigQuery and need to update the model daily. The team has data scientists who can write custom TensorFlow code but want to minimize infrastructure management. Which product should they use?

   • ABigQuery ML to train a matrix factorization model.
   • BVertex AI Pipelines to orchestrate custom training and deployment.
   • CVertex AI AutoML Tables to train a recommendation model.
   • DDeploy a pre-trained LLM to generate recommendations from article text.
   Reveal Solution Discussion Select an answer first