D2MR: A Framework for Recommending Drift Detection Method via Data Characteristic-based Rule Mapping in Machine Learning System
Okjoo Choi · Journal of Software Assessment and Valuation · 2025
Machine learning systems often suffer from performance degradation and reduced model reliability due to data distribution shifts (drift) that occur over time.Although various drift detection methods have been studied to address this issue, selecting a method that is well-suited to the characteristics of real-world datasets remains challenging.This study proposes the D 2 MR(Drift Detection Method Recommender) framework, which automatically recommends suitable detection methods based on dataset meta-features such as label availability, data type, size, dimensionality, distance computability, and distribution type.The proposed framework systematically maps data characteristics to detection methods and can be effectively integrated into drift monitoring modules of machine learning systems.Finally, the effectiveness of D 2 MR is validated through case studies using the CIFAR-10 image dataset and the UCI Wine Quality dataset.