A Unified Framework for Anomaly and Concept Drift Detection with Explainability

Concept drifts and anomalies can exist within real-world data streams. Identifying such events is of crucial importance as they can cause the deterioration of the performance of machine learning algorithms. Additionally, the integrity of the model's decision and the lack of labeled information have further complicated the data streams' analysis. Hence, this thesis studies the approaches for the individual research field of anomaly detection, concept drift detection, interpreting a model's decision process, and estimating the required information resources. A unified framework incorporating the four research fields is developed for a complete analysis of the dynamic data stream.