Section: New Results

Axis 2: Change-point detection by means of reproducing kernels

Participant : Alain Celisse.

Classical offline change-point detection approaches are limited to detecting changes arising in the mean and/or variance of the distribution along the time. Detecting changes in other moments of the distribution is possible, but at the price of stronger (unrealistic) distributional assumptions which are likely to be violated.

Reproducing kernels are a means to detect changes arising in any moments of the distribution along the time, which are not limited to the mean or the variance. One of the main contributions of this work is to provide a theoretically grounded model selection strategy allowing us to detection multiple changes. From additional extensive simulation experiments, it clearly arises that the so-called KCP approach outperforms numerous state-of-the art change-points detection procedures such as E-divisive, PELT, ...