In data streams, such as stock market transactions, concept drift occurs when the relationship between the input data and target variable changes. For its different categories, the reader may refer to [1]. This papers aims to provide a comprehensive framework for the quantitative analysis of drift. The framework relates different types of concept drift, discusses distinctions among them, provides a taxonomy, and tries to provide foundations for their more effective and efficient study.

The focus of the paper is redefining the concept drift categories over evolving data streams using the Hellinger distance. It provides a good discussion about the existing concept drift categories and then provides quantitative measures for recognizing concept drift types. The authors discuss their point of view on the behavior of online classifiers on some synthetic data with nonstationary distribution.

The problem is highly interesting and I like the idea for defining concept drift types in a formal way rather than an intuitive definition. As mentioned in the paper, there is a gap in the literature on recognizing different categories of concept drifts and acting accordingly. The paper requires careful reading and good preparation. Researchers will find it beneficial.