A retail forecasting model has run in production on Vertex AI for three months. The serving features come from a streaming pipeline, while the model was trained on a batch snapshot of the same source. Latency-sensitive predictions have grown steadily less accurate, but the live input feature distributions still match the training data when checked against the original training baseline. Which Model Monitoring signal is the team most likely observing, and what does it indicate?
- AFeature attribution drift, indicating that the relative importance the model places on each feature has shifted even though raw inputs look stable.
- BTraining-serving skew, indicating that the feature values produced at serving time differ from those produced by the training pipeline.
- CConcept drift, indicating that the statistical relationship between the input features and the target has changed over time. Correct
- DData drift, indicating that the distribution of the incoming serving features has moved away from the training distribution.
Why A is wrong: This is tempting because attribution drift can occur while inputs look stable, but attribution drift reflects changes in how features contribute to predictions, not a change in the input-to-target relationship that degrades accuracy with unchanged inputs and unchanged outputs.
Why B is wrong: Skew is plausible given the batch-versus-streaming pipeline split, but the stem states serving feature distributions still match the training baseline, so the inputs are not diverging between training and serving.
Why C is correct: Inputs match the training baseline yet accuracy falls, so the mapping from features to target has changed, which is the definition of concept drift and explains degradation without input distribution change.
Why D is wrong: Data drift is a natural first guess for falling accuracy, but it requires the serving input distribution to shift, and the stem explicitly says inputs still match the training baseline, ruling it out.