Question

A data science team trains a BigQuery ML model to predict customer churn. During training they compute several engineered features in the SELECT statement: a log transform of monthly_spend, a ratio of support_tickets to tenure_months, and a bucketized age. When the model is deployed and analysts run ML.PREDICT with raw customer rows, predictions are wildly inaccurate compared to evaluation results. The team wants the preprocessing to be applied automatically and identically at both training and prediction time with minimal maintenance. What should they do?

Accepted Answer

Wrap all feature engineering expressions inside a TRANSFORM clause in the CREATE MODEL statement so BigQuery ML applies them automatically during training and prediction.

Answer

Create a scheduled query that materializes the engineered features into a separate table and point ML.PREDICT at that table.

Answer

Document the preprocessing SQL and require analysts to reproduce the same expressions in every ML.PREDICT query.

Answer

Export the model to Vertex AI and rebuild the preprocessing as a separate Vertex AI pipeline component.