Question

A data science team is building a BigQuery ML logistic regression model to predict customer churn. One key input is customer_age, a continuous integer, and another is annual_income, which ranges from a few thousand to several million and is heavily right-skewed. During evaluation the model underperforms, and the team suspects the raw scale and skew of these numeric features are hurting training. They want the preprocessing logic to be applied consistently and automatically at both training and prediction time, without maintaining a separate transformation pipeline. What should they do?

Accepted Answer

Define a TRANSFORM clause in the CREATE MODEL statement that applies ML.BUCKETIZE to customer_age and ML.STANDARD_SCALER (with log handling) to annual_income so the transformations are stored with the model.

Answer

Manually compute bucketed and scaled columns in a staging table before training, then remember to apply the same SQL logic in every prediction query.

Answer

Enable AUTO_CLASS_WEIGHTS in the OPTIONS clause so BigQuery ML automatically normalizes the numeric feature distributions during training.

Answer

Export the data to Vertex AI, run feature scaling in a custom container, and re-import the scaled features into BigQuery for training.