Task 3.1: Building models given the task considering cost, complexity, latency, and scalability.

• ABigQuery ML logistic regression, with ML.ADVANCED_WEIGHTS for p-values and ML.EXPLAIN_PREDICT per applicant

  Correct

  Linear models in BigQuery ML provide standard errors and p-values for their weights and local Shapley attributions for each prediction.

• BA BigQuery ML boosted tree model, with ML.EXPLAIN_PREDICT for each applicant and ML.GLOBAL_EXPLAIN for the regulator

  Incorrect

  Tree SHAP gives per-row attributions, but tree models have no standard errors or p-values for significance testing.

• CA BigQuery ML DNN classifier, sharing its ML.GLOBAL_EXPLAIN results with the regulator

  Incorrect

  Global feature importance ranks the inputs; it is not a significance test, and the regulator asked for one.

• DAn Agent Platform AutoML tabular model, with its sampled Shapley feature importance shown to the regulator

  Incorrect

  AutoML offers sampled Shapley importance, not standard errors or p-values for each input's effect.

BigQuery Explainable AI uses a different method for each model type. Linear and logistic regression support Shapley values through ML.EXPLAIN_PREDICT and ML.GLOBAL_EXPLAIN, plus standard errors and p-values for significance testing through ML.ADVANCED_WEIGHTS. Boosted trees and random forests use Tree SHAP and Gini importance, and AutoML Tables uses sampled Shapley.

• AThe Cloud Vision API's label detection feature

  Incorrect

  Pretrained label detection returns general labels, not the manufacturer's own seven defect classes.

• BA custom PyTorch training job on the labeled photos in serverless training

  Incorrect

  Custom training needs training code and ML skills, which the team does not have.

• CA BigQuery ML logistic regression model over an object table

  Incorrect

  Logistic regression in BigQuery ML trains on tabular features, not on the raw image pixels of the photos.

• DAgent Platform AutoML image classification

  Correct

  AutoML trains an image classification model from labeled images without training code and can deploy it to an endpoint.

AutoML on Agent Platform trains image classification and object detection models from labeled images with no code. Pretrained APIs such as Vision label detection only return general labels, and custom training is for teams that write their own training code.

• ADeploy the AutoML model to an Agent Platform endpoint and call it from the cameras over the site's link

  Incorrect

  An endpoint needs a network round trip for every image, which the site cannot provide.

• BRun batch inference on the images each night after they are copied to Cloud Storage

  Incorrect

  Nightly batch results arrive hours later, far outside the 50 ms requirement.

• CTrain an AutoML Edge model and export it as a TF Lite package for the camera devices

  Correct

  AutoML Edge models export as TF Lite, Edge TPU TF Lite or a container, so inference runs on the device with no network.

• DDeploy the model to a private endpoint that uses Private Service Connect

  Incorrect

  A private endpoint still requires connectivity to Google Cloud, which the site does not have.

AutoML Edge image models can be exported as a TF Lite package for edge or mobile devices, as Edge TPU TF Lite, or as a container. That lets inference run where the data is produced when there is no connectivity or the latency budget rules out a network call.

• AA BigQuery ML matrix factorization model trained on the purchase history of all users and refreshed with every night's new orders

  Incorrect

  Matrix factorization learns embeddings only for items with interactions, so new products cannot be recommended.

• BA Gemini model that scores every product against the user's profile on each page view

  Incorrect

  Scoring 30 million products per request with an LLM cannot fit a 50 ms budget or a sensible cost.

• CA k-means model that groups products by category and recommends the user's favorite cluster

  Incorrect

  Clusters are coarse, so recommendations are barely personalized and ignore similarity within a cluster.

• DA two-tower model whose item tower uses content features, with item embeddings served from an approximate nearest neighbor index

  Correct

  The item tower can embed items never seen in training, and ANN search over precomputed embeddings meets the latency budget.

Two-tower retrieval learns a query tower and a candidate tower. Because the candidate tower takes item features, it can compute embeddings for items not seen during training, which addresses cold start. Candidate embeddings are precomputed and served from a Vector Search index configured for approximate nearest neighbor search.

• AK-means clustering on the demand, temperature and promotions columns of each region

  Incorrect

  K-means groups similar days; it does not forecast demand over a horizon.

• BA matrix factorization model over regions and days

  Incorrect

  Matrix factorization is for recommendations from user and item interactions, not for forecasting.

• CARIMA_PLUS_XREG, with temperature and promotions as regressors

  Correct

  ARIMA_PLUS_XREG extends the ARIMA_PLUS pipeline with external regressors, so extra variables shape the forecast.

• DA PCA model that reduces the columns before forecasting

  Incorrect

  PCA reduces dimensions; it does not produce a forecast.

BigQuery ML forecasts with ARIMA_PLUS for a single variable, ARIMA_PLUS_XREG when you need to forecast with additional variables (external regressors), or the pretrained TimesFM model. ML.FORECAST returns the forecast for the horizon.

• AThe Tabular Workflow for End-to-End AutoML

  Correct

  This generally available workflow exposes each step of AutoML, including architecture search, hardware and ensemble size.

• BStandard AutoML tabular training with a larger training budget

  Incorrect

  More budget gives AutoML more time but no control over search space, hardware or ensemble size.

• CThe Feature Transform Engine run on its own

  Incorrect

  Feature Transform Engine handles feature selection and transforms, and it is in Preview.

• DA BigQuery ML boosted tree model with default options

  Incorrect

  A boosted tree model is a different model type, not AutoML with more control over its steps.

Tabular Workflow for End-to-End AutoML is a managed Agent Platform Pipelines workflow that lets you choose what to control and what to automate at each step: data splitting, feature engineering, architecture search, model training and ensembling, including hardware and ensemble size. It is generally available, while Feature Transform Engine and the forecasting workflow are in Preview.

• ADynamic inference on every page load from an online endpoint that autoscales

  Incorrect

  On-demand prediction adds latency and cost for results that do not change during the evening, and they cannot be checked in advance.

• BCalling a large language model for each user at page load

  Incorrect

  An LLM call per page load is slow and costly for a fixed nightly list.

• CRunning the model on each user's device after the page loads

  Incorrect

  On-device inference still computes at page load and cannot be reviewed beforehand.

• DStatic inference that precomputes and caches each user's row in advance

  Correct

  Precomputed predictions are served instantly and can be verified before anyone sees them.

Static inference generates predictions in advance and caches them, which suits a known set of inputs when prediction speed, cost and the ability to verify predictions matter. Dynamic inference predicts on demand and suits diverse inputs that cannot be enumerated in advance.

• ARetrain the model as a BigQuery ML DNN classifier with default options on the same data

  Incorrect

  Retraining discards the team's validated PyTorch model and its architecture.

• BConvert the model to ONNX and import it with CREATE MODEL in BigQuery ML

  Correct

  BigQuery ML imports ONNX models, such as those from PyTorch, and runs prediction in BigQuery with ML.PREDICT.

• CDeploy the model to an endpoint and call it from a remote function

  Incorrect

  This adds an endpoint and per-row calls, which the team wants to avoid.

• DExport the table to Cloud Storage and score it in a notebook

  Incorrect

  Exporting and scoring outside BigQuery moves the data and leaves SQL.

BigQuery ML can import models trained outside BigQuery, including ONNX (from frameworks such as PyTorch and scikit-learn), TensorFlow, TensorFlow Lite and XGBoost, and run prediction in BigQuery. The CREATE MODEL statement for an imported model processes no bytes.