From Drift Detection to Diagnosis: An Explainable Framework for Deployed Machine Learning

Abstract

Machine learning (ML) models are increasingly deployed in real-world environments where data characteristics, user behaviour, and operating conditions may change over time. Such changes can lead to data or concept drift and may gradually reduce model reliability. Existing monitoring approaches can identify distributional changes or performance degradation, but they often provide limited information about why the changes occur. Similarly, Explainable Artificial Intelligence (XAI) methods can explain model predictions but are frequently treated separately from continuous drift monitoring. This paper proposes a unified drift-aware framework for explainable diagnostics in deployed machine learning systems. The framework integrates data-drift detection, model-performance monitoring, explainability, and diagnostic decision-making into a continuous monitoring pipeline. Statistical drift detection methods are used to identify changes between reference and production data, while XAI techniques such as SHAP and LIME are used to identify important features associated with changes in model behaviour. The framework combines drift indicators and performance information to generate interpretable diagnostic reports and recommend suitable maintenance actions. The proposed approach aims to improve the transparency, reliability, and maintainability of deployed ML systems by connecting drift detection with actionable explanations. The framework can support proactive model monitoring and assist practitioners in deciding whether continued monitoring, investigation, recalibration, or model retraining is required. Keywords: Explainable AI,Machine Learning Deployment, Model Monitoring,Drift Detection