Дослідження методів виявлення аномалій та зміщення даних (DATA DRIFT) у системах машинного навчання

Abstract

This work is devoted to investigating methods for anomaly and data drift detection in machine learning systems. The nature of data drift and anomalies, including covariate shift and concept drift, were considered. An analysis of ensemble methods for outlier detection and statistical metrics for continuous data drift monitoring is carried out. It has been found that combining these approaches within the MLOps architecture forms a reliable protective layer, preventing the silent degradation of intelligent systems in a changing real-world environment