From XAI to MLOps: Explainable Concept Drift Detection with Profile Drift Detection

Concept drift—where the relationship between input data and predictions fundamentally changes—is harder to catch than simple performance drops, and traditional monitoring methods that only track accuracy or data distributions may miss these critical shifts until significant damage is done.

The proposed Profile Drift Detection method uses explainability tools (Partial Dependence Profiles) to not only detect when drift occurs, but also help practitioners understand why it happened, moving beyond black-box alerts to actionable insights about which relationships in the data have changed.

This approach is computationally efficient enough for real-time monitoring, addressing a practical barrier to implementing drift detection in production systems where continuous model surveillance is necessary but resource-intensive.

The method successfully balances competing priorities in model monitoring: it maintains predictive performance while avoiding false alarms that could trigger unnecessary retraining, a key challenge in responsible AI deployment where over-reactive systems waste resources and under-reactive ones fail users.

By connecting explainability (XAI) with operational practices (MLOps), this work demonstrates that responsible AI requires both understanding model behavior and maintaining robust systems, rather than treating these as separate concerns.