A data analyst has trained a classification model named customer_churn_model using BigQuery ML to predict customer attrition. The analyst now needs to generate batch predictions on a new dataset of current subscribers stored in a table named active_customers and output the predicted labels alongside customer identifiers for a marketing campaign. Which BigQuery ML function should the analyst use to execute these predictions?

ML.PREDICT: ### **What ML.PREDICT is:** `ML.PREDICT` is the standard **BigQuery ML function** used to generate predictions, classifications, or cluster assignments using a previously trained machine learning model. It allows data practitioners to run batch inference directly inside BigQuery using standard SQL syntax without exporting data or managing external prediction microservices. ### **How it meets the requirements:** - **Batch inference execution (✓):** `ML.PREDICT` takes the trained model name (`customer_churn_model`) and an input table or subquery (`active_customers`) to generate predicted labels for each record in parallel. - **Direct SQL integration (✓):** The output of `ML.PREDICT` includes all original input columns alongside new prediction columns (such as `predicted_` and `predicted__probs`), making it easy to filter, aggregate, or join directly with downstream marketing and reporting tables. - **Serverless scaling (✓):** It leverages BigQuery's distributed compute engine to perform predictions over millions of records in seconds without provisioning infrastructure. ### **Key advantages and capabilities:** - **Seamless integration:** Operates within standard SQL queries, enabling predictions to be written directly to new tables or visualized in Looker. - **Preserves input columns:** Automatically passes through all feature columns and identifying keys (like `customer_id`) from the input data alongside prediction outputs. - **Probability outputs:** For classification models, it returns prediction probabilities for each potential class, enabling granular thresholding and targeting. - **Broad model support:** Works across multiple model types, including logistic regression, linear regression, boosted trees, random forests, deep neural networks (DNNs), and imported TensorFlow or ONNX models. ### **Why this is optimal:** `ML.PREDICT` is specifically designed for supervised and unsupervised batch inference on tabular datasets. It minimizes operational complexity and latency by executing predictive scoring directly where the data resides.