Concept drift estimation with graphical models

This paper deals with the issue of concept-drift in machine learning context high dimensional problems. In contrast to previous concept drift detection methods, this application does not depend on model use for a specific target variable, but rather, it attempts assess as an independent characteristic evolution dataset. major achievement enables data be tested presence drift, independently problem at hand. is extremely useful when same dataset utilized different classifications simultaneously, often case business environment. Moreover, unlike approaches, method require re-testing each new model; strategy which could prove expensive computational terms. The fundamental intention work make graphical models elicit visible structure and represent network. Specifically, we investigate how evolves by looking creation links, disappearance existing ones, time periods. We perform task four steps. compute adjacency matrix graph period, apply function that maps possible state over into transition matrix. information produce metric estimate data. Eventually, evaluate both three real-world datasets synthetic one.