Wild Binary Search: New Paradigm for Interpretable Nonparametric Curve Estimation?

• Activity Number: 161
• Type: Topic Contributed
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 12:20 PM
• Sponsor: IMS
• Author(s): Piotr Fryzlewicz*
• Companies: London School of Economics
• Keywords: multiscale methods ; change-point detection ; cusp detection ; isotone regression ; randomized algorithms ; hierarchical methods

Abstract

We propose a new generic procedure for nonparametric curve estimation and for the detection of general features (e.g. cusps, locally isotonic regions) in noisy curves. The procedure consists of fitting the best single feature to the data in each of a number of randomly chosen contiguous subsets of the domain, and then aggregating the results in a fast hierarchical way. This is a generalization of the recently proposed Wild Binary Segmentation methodology for change-point detection (Fryzlewicz, 2014, Ann. Stat.).