Professional Machine Learning Engineer

Question

You work at a bank and have a custom tabular ML model supplied by the bank’s vendor. Its training data is unavailable because it is sensitive. The model is packaged in a Vertex AI Model serving container that accepts one string per prediction instance, with feature values in each string separated by commas.

You want to deploy this model to production for online predictions and, with minimal effort, monitor how the feature distribution changes over time. What should you do?

Explanation

Feature drift detects changes in feature distributions in production data over time and does not require the unavailable training dataset as a reference. For a custom model receiving CSV-string inputs, an analysis instance schema specifies the payload format, feature names, types, and feature order, allowing Model Monitoring to parse those inputs without changing the serving container. Feature skew instead compares serving data with training data, so it is unsuitable when the training data is unavailable.

Learn more

Provide schemas to Vertex AI Model Monitoring

Professional Machine Learning Engineer

AstroTutor

Notes

Readiness

QuestionQ56

Community Discussion