DcDM: A Pre-training Data Evaluation Framework for Proactive Drift Prevention in Machine Learning

Abstract

The performance of machine learning (ML) systems often deteriorates over time owing to data drift, which is typically caused by changes in data quality or distribution. Such degradation in deployment environments can result in inaccurate predictions and reduced system reliability. Conventional drift detection approaches have largely focused on retraining ML models after performance degradation has occurred. However, because the root causes of drift often originate from the data itself, a data-centric approach is needed to address the problem at its source. This study proposes a systematic, data-centric drift management (DcDM) framework that integrates domain-specific rule validation, data quality assessment, and statistical drift analysis before model training. By first verifying semantic constraints and then assessing data quality and distributional stability, DcDM enables early identification of potential drift hazards and prevents low-quality or semantically invalid data from entering the training pipeline. We evaluate the proposed framework on three datasets across different modalities: Electricity Load Diagrams, CIFAR-10, and VisDA-2017. The experimental results show consistent improvements of approximately 6% to 12.5% in both accuracy and F1-score. Additional ablation studies, baseline comparisons, and statistical significance tests further demonstrate the robustness and effectiveness of the proposed approach.