همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

A Multi-Dimensional Taxonomy and Practical Framework for MLOps Monitoring

نویسندگان :

Tina Hafezi¹ Sara Pour Faraj² Seyed Mohammad Amin Alemohammad³ Niloofar Mirzaei Chahardeh⁴

1- دانشگاه ملی مهارت 2- دانشگاه ملی مهارت 3- دانشگاه آزاد اسلامی واحد تهران مرکزی 4- دانشگاه آزاد اسلامی واحد علوم و تحقیقات

کلمات کلیدی :

mlops،machine learning monitoring،data drift،concept drift،model metrics

چکیده :

Unlike conventional software systems, machine learning (ML) systems operate in dynamic environments that can contain non-stationary data, delayed labels, and continuously changing behavior. These differences make monitoring a fundamental competency within MLOps, providing the observability needed to maintain the reliability, stability, and long-term business value of ML systems. This paper proposes a three-dimensional taxonomy of ML monitoring metrics encompassing data quality constraints, statistical drift metrics, classical performance indicators, behavioral proxies, system health measures, and business performance metrics, including cost and value dimensions, to address the most common practical scenarios. Building upon this taxonomy, we introduce a practitioner-oriented framework for selecting, integrating, and operationalizing monitoring metrics across the ML lifecycle. The framework incorporates metrics aligned with stakeholder objectives and links technical and business indicators. It also integrates early-warning diagnostics, automated responses, and both label-dependent and label-independent monitoring. Furthermore, for maintaining resilience against data drift and concept drift, interpretability, human-in-the-loop evaluation, and automated retraining mechanisms are emphasized as essential elements. We also outline observability modules, model lineage tracking, automation triggers, and retraining logic that operationalize the proposed framework. The presented approach provides a unified foundation for designing robust monitoring systems that preserve model performance over extended periods, reduce operational costs, and strengthen the reliability and business alignment of production ML systems.