A dynamic routing algorithm of CapsNet for drift prognosis

In data stream mining, detecting significant changes in a data stream is called drift detection. Detecting drift before it starts is an important problem, but there is very limited research. We call it “drift prognosis”. Many existing drift detection methods only report a drift after it has occurred. This paper tackles this challenge, taking advantage of Capsule Networks (CapsNets), a recent deep-learning architecture. CapsNets can encapsulate the properties of features. We propose a novel dynamic routing algorithm for drift prognosis, named DR-DD, which can transform between capsule layers to capture subtle changes, indicating a potential drift. Compared to 11 drift detection methods in the literature, our DR-DD algorithm is the only one that can pre-diagnose a drift, before it occurs.