Task 1.1: Developing ML models using BigQuery ML or AutoML on Gemini Enterprise Agent Platform.

• AA boosted tree regression model in BigQuery ML, trained on lag, trend and holiday features that the analysts build in SQL for each store and product

  Incorrect

  A regression model needs the lag, trend and holiday features to be engineered by hand, which ARIMA_PLUS handles itself for each series.

• BAn Agent Platform AutoML forecasting model trained from a Python notebook, with its forecasts exported to BigQuery after each run

  Incorrect

  AutoML forecasting can do this, but it moves the work out of SQL into a notebook workflow the analysts do not use.

• CA separate ARIMA_PLUS model per combination, created by a scheduled script that loops over the pairs and calls ML.FORECAST on each

  Incorrect

  This gives 18,000 model definitions to maintain, which the team ruled out; TIME_SERIES_ID_COL fits them all in one statement.

• DOne ARIMA_PLUS model with TIME_SERIES_ID_COL on store and product and HOLIDAY_REGION set, with every series read through ML.FORECAST

  Correct

  A single CREATE MODEL statement fits a separate time series pipeline for every ID, models holiday effects for the region, and ML.FORECAST returns the horizon as a table the analysts can query.

The ARIMA_PLUS model in BigQuery ML handles seasonality and holidays automatically. With TIME_SERIES_ID_COL, one query trains and forecasts up to 100,000,000 time series in parallel, and HOLIDAY_REGION adds the holiday effects of the chosen region. ML.FORECAST then returns the forecast for the horizon as rows in BigQuery.

• APrecision at Recall, with its recall value set to the required 0.9

  Correct

  This objective optimizes precision at a specific recall value, which is exactly the fixed-recall, fewest-false-alarms requirement.

• BAUC ROC, the default objective for binary classification

  Incorrect

  AUC ROC distinguishes the classes over all thresholds; it does not hold recall at the required 90%.

• CAUC PR, to focus the model on the less common class

  Incorrect

  AUC PR helps with the rare class across thresholds, but it does not optimize for the stated recall operating point.

• DLog loss, to keep the predicted probabilities accurate

  Incorrect

  Log loss targets well-calibrated probabilities, not the precision at a given recall.

AutoML classification lets you change the default optimization objective. Precision at Recall (maximize-precision-at-recall) optimizes precision at a recall value you set, and Recall at Precision does the reverse. AUC ROC is the default for binary classification, AUC PR favors the less common class, and log loss keeps probabilities accurate.

• ARetrain the model with AUTO_CLASS_WEIGHTS set to TRUE

  Incorrect

  Class weights change how the model is trained, and the team wants the model left as it is.

• BPass STRUCT(0.8 AS threshold) as the last argument of ML.PREDICT

  Correct

  ML.PREDICT accepts a custom threshold for binary classification; scores above it become positive predictions.

• CRun ML.EVALUATE with a threshold of 0.8 before each scoring run

  Incorrect

  ML.EVALUATE reports metrics at a threshold; it does not change the labels that ML.PREDICT returns.

• DKeep the rows whose predicted label is 1 in the default ML.PREDICT output

  Incorrect

  The default output uses a 0.5 cutoff, so it marks far more customers as positive than the team can call.

For binary classification models, ML.PREDICT takes an optional THRESHOLD argument (default 0.5) in a STRUCT, for example STRUCT(0.55 AS threshold). Predictions above the threshold are positive. This changes the operating point without retraining the model.

• ACreate a remote model over the base Gemini endpoint and add twenty example tickets to every prompt that the analysts send through AI.GENERATE_TEXT

  Incorrect

  Few-shot prompting can help style, but it does not use the 40,000 examples and adds tokens to every request; it is not tuning.

• BTune a remote Gemini model with CREATE MODEL AS SELECT on the ticket columns, check it with ML.EVALUATE, then use AI.GENERATE_TEXT

  Correct

  The CREATE MODEL statement for tuning runs supervised fine-tuning on the selected rows, and the tuned remote model is evaluated and used from SQL.

• CExport the tickets to Cloud Storage as JSONL, start a supervised tuning job from an Agent Platform notebook and call the tuned model from Python

  Incorrect

  Tuning in Agent Platform works, but it moves the work out of SQL, which the analytics team ruled out.

• DTrain a BigQuery ML boosted tree model on the messages and use it to pick one of the existing replies

  Incorrect

  A classifier can predict the category but cannot write new replies in the experts' style.

BigQuery ML can fine-tune a Gemini model: CREATE MODEL with REMOTE WITH CONNECTION, the ENDPOINT of a supported model and an AS SELECT of prompt and label columns runs supervised tuning. You then evaluate the tuned model with ML.EVALUATE and generate text with AI.GENERATE_TEXT or AI.GENERATE_TABLE, all in SQL.

• AML.FORECAST on the model, comparing each reading with the forecast by hand

  Incorrect

  ML.FORECAST predicts future values; flagging past readings would need extra logic that ML.DETECT_ANOMALIES already provides.

• BML.DETECT_ANOMALIES on the ARIMA_PLUS model

  Correct

  ML.DETECT_ANOMALIES supports ARIMA_PLUS models for time series data and returns an anomaly flag and probability for each point.

• CML.PREDICT on a k-means model trained on all pumps together

  Incorrect

  K-means suits independent data points, not each pump's time pattern, and the team already has the time series model.

• DML.EVALUATE on the ARIMA_PLUS model for each pump

  Incorrect

  ML.EVALUATE reports how well the model fits; it does not flag individual readings.

ML.DETECT_ANOMALIES performs anomaly detection in BigQuery ML. For time series data it works with ARIMA_PLUS time series models; for independent and identically distributed data it works with k-means, PCA and autoencoder models.

• AA NUM_TRIALS option, with HPARAM_RANGE for the two settings

  Correct

  Setting NUM_TRIALS turns on hyperparameter tuning, and the ranges or candidates define the search space for each hyperparameter.

• BA TRANSFORM clause that lists the two hyperparameters

  Incorrect

  TRANSFORM defines feature preprocessing, not a search over model settings.

• CA DATA_SPLIT_METHOD of RANDOM, which reshuffles the rows on every training run

  Incorrect

  The split method decides how rows are divided; it does not try other hyperparameter values.

• DA separate CREATE MODEL statement for every combination of values

  Incorrect

  Separate models are the manual search the analyst wants to avoid.

BigQuery ML runs hyperparameter tuning when you specify NUM_TRIALS in CREATE MODEL. Search spaces are given with HPARAM_RANGE or HPARAM_CANDIDATES, and tuning automatically splits the input into training, evaluation and test sets.

• AA random split with a larger test fraction of the rows held out

  Incorrect

  A bigger random test set still mixes future rows into training, so it overstates accuracy the same way.

• BA stratified random split on the store column

  Incorrect

  Agent Platform does not perform stratified sampling, and any split of the rows by store still lets the model train on rows later than the ones it is tested on.

• CA chronological split on the transaction date column

  Correct

  A chronological split puts the most recent data in the test set, which matches how the model is used on future data.

• DNo split, with every row used for training

  Incorrect

  Without held-out data there is no honest measure of performance at all.

For tabular training, Agent Platform supports random, manual and chronological splits. When the data is time-sensitive, use a chronological split on the time column, or a manual split that puts the most recent data in the test set, so that evaluation reflects performance on future data.

• AThe optimization objective of the model

  Incorrect

  The objective decides what the model optimizes, not how long it trains.

• BA weight column on the training data

  Incorrect

  A weight column changes how much each row counts during training.

• CThe chronological data split

  Incorrect

  The split decides which rows train and test the model; it does not cap the cost.

• DThe training budget in node hours

  Correct

  The training budget sets the maximum time the model trains, in node hours, which caps the training cost.

When you train an AutoML tabular model, you set the maximum number of node hours it can train for (trainBudgetMilliNodeHours in the API), which caps the training cost. Early stopping can end training sooner.