Kernel Change-Point Detection

This algorithm takes pre-processed (interpolated to a fixed time grid, normalized, etc.) and offline (not streaming) multidimensional time series data and runs linear kernel anomaly/change-point detection on it in order to output a list of likely anomaly/change-points in time. The method is based on results in Arlot et al. (2019).

We have started to develop automated post-hoc visualization tools in order to provide intuitive explicability in the output results in order to attribute a ranked (in terms of importance) list of individual time series for each detected anomaly/change-point. This is because it is otherwise difficult to know 'why' a change-point was detected at a certain point if there are hundreds or thousands of individual concurrent time series. The latter tool is semi-dependent of the anomaly/change-point detection step.