Which approach helps address both data drift and concept drift in a deployed AI system?

• A. Increasing Batch Size
• B. Monitoring Input Distributions And Model Performance
• C. Reducing Dataset Size
• D. Removing Logs

Answer: (D)

To address both data drift and concept drift in a deployed AI system, you need visibility into how inputs evolve over time and how the model’s outputs behave. Monitoring the distributions of input features over time reveals data drift—when the statistical properties of what the model sees change from what it was trained on. At the same time, watching model performance metrics (such as accuracy, precision/recall, or domain-specific KPIs) over time helps you spot concept drift—when the relationship between inputs and the target changes, causing predictions to degrade even if inputs haven’t changed dramatically. Together, these monitoring signals let you detect when retraining, feature updates, or threshold recalibration are needed, and they support a proactive maintenance loop for the deployed system.

In contrast, simply increasing batch size doesn’t provide visibility into drift; reducing dataset size reduces information available to detect changes; removing logs eliminates the very data you need to monitor drift and performance.