머신러닝 시스템에서 데이터 품질 평가 기반의 데이터 중심 드리프트 관리 방법론 = A Data-centric Drift Management Methodology based on Data Quality Assessment in Machine Learning System

다국어 초록 (Multilingual Abstract)

Machine learning (ML) systems often suffer from degradation in prediction accuracy and reliability due to data drift that occurs over time. Conventional approaches primarily focus on restoring performance through model retraining; however, it is essential to recognize that the root cause of drift lies in the data itself. In this study, we propose the Q-DcDM framework, which integrates data requirement–driven quality assessment with drift analysis to proactively manage drift before the model training and inference stages. The proposed framework enables early detection of potential drift in the initial phase of the ML lifecycle, thereby preventing performance degradation. Experimental results on the UCI Wine Quality and CIFAR-10 datasets demonstrate that the proposed approach improves average model performance by 14.8% and 12.3%, respectively, while also significantly reducing the frequency of model retraining.