Title: Change-point regression robustified to smooth artifacts
Authors: Florian Pein - University of Cambridge (United Kingdom) [presenting]
Rajen D Shah - University of Cambridge (United Kingdom)
Paul Fearnhead - Lancaster University (United Kingdom)
Abstract: Piecewise constant regression is undeniable the most common form of change-point regression. While the assumption of a piecewise constant signal is reasonable in many applications, there are import examples, for instance genome sequencing to detect copy number variations or ion channel recordings (experiments to measure the conductance of a single ion channel over time), where such an assumption is at least questionable. Instead, such experiments can be better modelled by a piecewise constant function plus a smooth function. Existing methods are not using such a model explicitly for detecting change-points. Contrarily, they assume a piecewise smooth signal, either explicitly or a decomposition into piecewise constant plus smooth functions is stated but not used for detecting change-points, instead only the smooth function is reestimated in a second step. To use the decomposition explicitly, a modified fused lasso combined with smoothing techniques is proposed. Simulations show that this leads often to a better detection power. Moreover, the new methodology is very flexible. Kernel regression, smoothing splines and other methodologies can be used for smoothing. Secondly, extensions to multivariate and to filtered datasets are straightforward. Both extensions will be used to analyse genome sequencing data and ion channel recordings, respectively.