This document discusses understanding change in versioned knowledge organization systems (KOS) on the web. It describes work being done in the CEDAR project to harmonize and publish historical Dutch census data as RDF data cubes. The document outlines challenges with concept drift (changes in meaning over time) in dynamic classifications and ontologies used in census data from different time periods. It proposes using machine learning techniques to predict where and when versioned KOS on the web will change based on analyzing patterns of change in past versions. Features related to structural changes and changes in class membership are discussed for use in the machine learning models.