Model-based explanations of concept drift

Concept drift refers to the phenomenon that distribution generating observed data changes over time. If is present, machine learning models can become inaccurate and need adjustment. While there do exist methods detect concept or adjust in presence of drift, question explaining i.e., describing potentially complex high dimensional change a human-understandable fashion, has hardly been considered so far. This problem importance since it enables an inspection most prominent characteristics how where manifests itself. Hence, human understanding increases acceptance life-long models. In this paper, we present novel technology characterizing terms characteristic spatial features based on various explanation techniques. To so, propose methodology reduce are trained suitable way extract relevant information regarding drift. way, large variety schemes available. Thus, method be selected for at hand. We outline potential approach demonstrate its usefulness several examples.