Understand the raw feature contract
For each input, record its type, unit, valid range, missing-value meaning, and availability at prediction time. A missing field can mean an unknown value, an unavailable source, or a meaningful absence; treating these as identical may hide important information. Inspect missingness and invalid values by time and relevant group. Define whether to reject, impute, add a missingness indicator, or use another approach based on the task and model.
A customer’s account age is absent because an upstream service timed out. Contrast this with an account that genuinely has no history. Define a safe fallback and one monitoring signal for each case.
Fit transformations on training data
Split examples first. Compute imputation values, scaling statistics, and category vocabularies from the training data, then apply those fitted values to validation, test, and serving data. For example, if the training mean is 10 and standard deviation is 2, a value of 14 has standardized value 2. Recomputing the mean on the test period lets information from that period influence evaluation. The exact handling of zero variance and missing values must be specified.
A numeric training feature has mean 10 and standard deviation 2. Standardize 8 and 14: the results are -1 and 2. Explain why using a mean computed from the future test period is inappropriate for an honest evaluation.
Plan for unseen categories and BigQuery ML behavior
A production category may not appear in training. Define and test an unknown-category strategy instead of silently assigning an arbitrary known label. BigQuery ML applies automatic preprocessing for supported models and also supports explicit transformations in the CREATE MODEL TRANSFORM clause. Its documentation states that supported analytic transformation statistics from that clause are reused during prediction; exact behavior depends on the function and model. Inspect transformed data where appropriate and verify current documentation.
Training contains device types “web” and “mobile”; serving receives “tablet.” Describe how you would test the configured encoder’s output, monitor unseen-category frequency, and decide whether retraining is warranted.
Test end-to-end parity
A shared transformation definition reduces duplication but does not prove the raw data, timing, units, or upstream joins are identical. Create representative examples and compare training and serving feature vectors after transformation, including missing, unseen, extreme, and boundary values. Version the transform with the model and feature contract. Monitor violations after deployment, and roll back or fall back when a mismatch makes predictions unsafe.
Training uses currency in dollars and serving receives cents. Specify one raw example and the expected transformed value, then describe a test that fails if the serving pipeline applies a different unit convention.
How should a scaler used for model evaluation obtain its mean and standard deviation?
Keep exploring
- BigQuery ML automatic preprocessing ↗
- Manual preprocessing and TRANSFORM ↗
- Numerical feature normalization ↗
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.