Unsupervised Concept Drift Detectors: A Survey

Concept drift mainly refers to the change of the current data distribution in the data streams due to the dynamic evolution of the external environment, which leads to the failure of machine learning or data mining models to show the desired effect. As a result, the forecast or decision model needs to be constantly updated. To find the appropriate time for model updating, a large number of studies have proposed methods for detecting concept drift, which can be divided into supervised detection methods and unsupervised detection methods. Because an unsupervised concept drift detector does not make strong assumptions about the availability of data annotations for real application scenarios, it has more robust availability and generality, and there is less summary work related to unsupervised detectors. Therefore, in this paper, we will summarize all the existing unsupervised concept drift detection methods and give a new classification basis to classify the unsupervised concept drift detection methods.