CapsNet-based drift detection in data stream mining

For data streams, drift detection methods warn and detect the changes in patterns over time. For example, in smart manufacturing, many data streams are generated from sensors that monitor the real-time operation of manufacturing. Drift detection can be used to discover if and how the operation status changes. At present, there have been three main approaches in drift detection: error rate-based, distribution-based, and hypothesis-based. However, these approaches bear an impractical limitation: delays due to the demand for computational time. In a large-scale and high-speed data stream, a time-efficient detector is vital. To address this, this paper proposes a CapsNet-based drift detection algorithm (CapsNet-DDM). Our experimental results and comparative studies have found that CapsNet-DDM demonstrates a distinguishing advantage on computational time, with no compromise on accuracy, F1 score, and effective drift detection rates.