Google Cloud Professional Machine Learning Engineer: Architecting low-code ML solutions Practice Questions

Options 1 and 2 are correct. BigQuery ML matrix factorization (option 1) is ideal for collaborative filtering on implicit feedback (purchase history) at scale. Vertex AI AutoML Tabular (option 2) excels when you have rich tabular features (user profile, product metadata) and a discrete target (next category), making it suitable for this multi-feature prediction task. Option 3 (Recommendations API) is a pretrained solution, but it is less flexible for this specific use case where custom features dominate the prediction signal. Option 4 (K-means) is unsupervised clustering and does not directly predict the next category. Option 5 (Vision API) is for image classification, not recommendations. Option 6 (custom TensorFlow) is high-code and not the 'most appropriate' given low-code options exist.

Options 2 and 3 are correct. Option 2 identifies distribution shift, a leading cause of model degradation in medical imaging—CT scanners vary by manufacturer, protocol, and hospital, and training on one site's images does not guarantee generalization to another. Option 3 addresses class imbalance in production; if training was 50-50 nodule/non-nodule but production is 95% non-nodules, the decision threshold may be miscalibrated, causing false positives on rare cases. Option 1 is incorrect; 92% accuracy suggests sufficient training data (AutoML Vision works with fewer than 100K images). Option 4 is incorrect; scheduled retraining every 30 days is not standard practice and would introduce latency. Option 5 is partially true but misleading; 92% accuracy does not prove hyperparameter optimization failed—adjusting the classification threshold is a post-training step, not a hyperparameter tuning issue.

Option 1 is correct because class balance and minimum examples per class are critical for AutoML performance. Vertex AI AutoML Text typically needs at least 100 labeled examples per class; imbalanced data degrades both model quality and the ability to evaluate minority classes. Option 2 is incorrect; AutoML Text supports far more than 8 classes (limits are typically 100+ classes). Option 3 is incorrect; AutoML Text works with varying text lengths; 500 characters is not a hard requirement. Option 4 is incorrect because BigQuery ML does support text classification via embedding-based models and text feature engineering, though AutoML Text may be more user-friendly.

Option 2 is correct because it uses Translation API to normalize all contracts to a single language (English), then leverages Natural Language API's entity extraction and syntax analysis to identify structured information (names via PERSON entities, dates via DATE entities, amounts via analysis). This requires minimal custom logic. Option 1 is incorrect because Document AI custom extractors require substantial training and annotation effort per language—not minimal code. Option 3 is incorrect because Natural Language API does not reliably extract custom fields like 'contract amounts' without additional processing; translation to English first ensures consistency. Option 4 is incorrect because Speech-to-Text is for audio input, and the problem states contracts (implying documents).

Option 3 is correct because BigQuery ML provides the native ML.EXPLAIN_PREDICT function, which computes per-prediction feature attributions without retraining or exporting—this directly answers the stakeholder need. Option 1 is incorrect; while LINEAR_REG coefficients can be inspected, they provide global feature importance, not per-prediction explanations required by the question. Option 2 is incorrect because Vertex AI Explainable AI (Integrated Gradients, SHAP) is designed for models registered in Vertex AI Model Registry, and BigQuery ML models are not natively integrated with this (though workarounds exist, they are not the primary approach). Option 4 is incorrect because retraining was explicitly excluded in the question.