Gradual Drift Detection by Computing Outlier in Data Stream using Z-score

In data streaming scenario, data are created successively. And it is very common that the concepts of data may change at any positions in data stream. The properties of the target variables may shift between the static training data to real-world dynamic data. For changing the concepts, the learning performances are affected seriously. In addition, due to storage limitations, it is tough for an online learner to predict original labels of streaming data because of the concept drift as well as the location of drifting for data streaming. Therefore, a method that can detect the exact positions of drifts and quickly able to adapt changes for improving the learning performance is needed. As a solution, this paper presents a Gradual Drift Detection method for Data Stream using Z-score (DDZ) method. This DDZ approach uses Z-score to detect the stable positions of drifts in data stream. The proposed DDZ method adapts the changes very fast. As a result, DDZ is able to predict correct labels of instances and provide better outcomes. Four synthetic data stream generators with abrupt and gradual drifting and two real-world datasets are considered for evaluating the performance of the proposed method. The performance of the proposed DDZ method attains higher accuracy, recall and F-Score as compare to some existing drift detectors.