Computing Divergences between High Dimensional Graphical Models

When the underlying distribution of a stream of incoming data changes, existing models trained on previous data may degrade in performance. Known as concept drift, one way to characterise these changes, is by measuring the magnitude at which these distributions change. However, in high dimensions, do so directly is practically impossible. Therefore, this thesis develops a method that can measure these high-dimensional distributional changes, by exploiting existing relationships between the variables in the distribution. This thesis then shows that this method provides a whole new set of possibilities with regards to visualising, analysing, and adapting to concept drift.