In data stream environments, the underlying data distribution often changes over time, a phenomenon known as concept drift. Drift detection methods (DDs) are designed to identify these changes and trigger model adaptation, but their performance strongly depends on internal hyperparameters that control sensitivity, detection delay, and stability, which directly affect their quality and computational performance. While extensive benchmarks have demonstrated the importance of configuration choices, systematic approaches for tuning unsupervised drift detectors
remain largely unexplored.

This thesis proposes a metadata-driven recommendation system that predicts suitable hyperparameter configurations for unsupervised drift detectors based on dataset characteristics. Using a large-scale benchmark dataset as the empirical foundation, a meta-dataset is constructed that links dataset-level features, detector parameters, and performance results.

The thesis aims to investigate whether efficient drift detector configurations can be inferred from metadata, exploring the potential for automated and computationally efficient pre-deployment tuning of unsupervised drift detectors. Beyond enhancing configuration selection, the research seeks to establish a foundation for future meta-learning approaches in adaptive and unsupervised drift detection.