Question

You deployed a customer-support text classifier six months ago on Vertex AI. Its accuracy on delayed ground-truth labels has slowly declined even though input-token distributions (length, language mix) appear stable in your monitoring dashboards. Investigation shows customers now describe new product features using terminology that did not exist in the training corpus. What is the most accurate characterization of this problem and the appropriate response?

Accepted Answer

This is concept drift where the relationship between inputs and labels has changed; retrain the model on recent labeled data that includes the new terminology.

Answer

This is training-serving skew caused by a preprocessing mismatch; align the serving tokenizer with the training tokenizer to restore accuracy.

Answer

This is data (input feature) drift only; add outlier clipping to the input pipeline and no retraining is needed.

Answer

This is a latency/serving degradation issue; increase endpoint replicas and enable request-response logging to recover accuracy.