Linear Semiparametric Estimation Procedure for the Segmentation of Multiple Series Based on a Lasso Strategy

• Activity Number: 131
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Cristian Meza*+ and Karine Bertin and Xavier Collilieux and Emilie Lebarbier
• Companies: Universidad de Valparaiso and Universidad de Valparaiso and IGN-Université Paris Diderot and AgroParisTech
• Keywords: Lasso estimator ; Semi-parametric model ; Joint segmentation

Abstract

In this work, we propose a new semi-parametric approach to the joint segmentation of multiple series. We use an iterative procedure based on Dynamic Programming for the segmentation part and Lasso estimators for the nonparametric bias part given a fixed number of segment. Then we choose the number of segments by a model selection criteria. The performance of our method is assessed using simulated data and we compare it to the methods developed in Picard et al (2011) where the bias part is viewed as a fixed effect or estimated by spline. Indeed, in comparison with these last methods, our Lasso procedure, based on the dictionary approach, allows to both estimate smooth functions and functions with local irregularity, which permits a more flexible modelization of the biases. As a consequence, we obtain a better estimation of the nonparametric part and improvements in the segmentation. We also apply our procedure to geodetic coordinate series in which it is important to correct for abrupt changes and periodic biases in order to determine accurate averaged positions and velocities or to study long period variations to infer the impact of climate change.