DDAL: A Drift Detection Active Learning Mechanism

Existing active algorithms cannot handle multiple concept drift patterns, discard historical data during retraining, and suffer from imbalanced training data. These limitations result in online algorithms being unable to maintain continuous high-precision detection. If data annotators want to know about the occurrence of concept drift through algorithms, they need to accurately identify concept drift in the data annotation phase. However, in practical scenarios, the data annotation system cannot accurately perceive concept drift. Therefore, this paper proposes a method called DDAL (Drift Detection Active Learning), which has the ability to perceive concept drift. DDAL manages data queues using time slices and incorporates concept drift detection, bringing forward the concept drift perception of online learning algorithms to the data annotation phase. This makes the model training more accurate and reduces the response time to concept drift. Through experiments and theoretical analysis, data annotated using the DDAL algorithm has higher accuracy compared to existing semi-supervised and unsupervised annotation methods.