In this paper, we propose DriftLens, an unsupervised learning-based real-time concept drift detection and characterization framework to address the problem of concept drift in real-world machine learning models that require continuous concept drift detection and monitoring. DriftLens overcomes the limitations of inefficient and inaccurate existing methods by leveraging the distributed distance in deep learning representations, enabling efficient and accurate detection. In addition, it characterizes and explains concept drift by analyzing the impact on each label. Evaluation results on various classifiers and data types show that DriftLens outperforms existing methods in 15/17 use cases, generates drift curves that are at least 5 times faster, has a high correlation (>0.85) with the actual drift, and effectively identifies and explains representative drift samples.
