A Concept Drift Detection Algorithm based on Fuzzy Marginal Density

With the continuous emergence of data, concept drift in data streams is becoming more and more common. In the past, concept drift detection was regarded as a task based on supervised learning, but it is of practical significance to study concept drift in semi-supervised or unsupervised learning since it is difficult to get all the labels of the data streams in real time. Existing algorithms based on semi-supervised learning to detect concept drift show good performance, but there is still room for improvement in terms of detection delay and false alarm rate. In this paper, we propose an algorithm named as Fuzzy Margin Density Drift Detection suitable for semi-supervised learning. This method explores the membership function of the fuzzy marginal dataset to more accurately describe and quantify the classification confidence of samples in the data stream, which takes full advantage of the classification confidence of each samples. This method is more accurate for concept drift detection, and can avoid the false alarm in some degree. We verified the effectiveness of the proposed algorithm through experiments on synthetic and real data set.