A1335
Title: Changepoint detection as model selection: A general framework and L0 approximation
Authors: Xueheng Shi - University of Nebraska-Lincoln (United States) [presenting]
Abstract: Changepoint detection is reframed as a model selection problem and introduces a unified framework that fundamentally broadens its scope. The framework accommodates a wide range of model structures including seasonal patterns, linear and quadratic trends, and autoregressive dependence, which allows changepoint analysis to move beyond restrictive, simplified settings. At the core of this framework is the iteratively reweighted fused lasso (IRFL), an efficient and scalable approximation to the L0 optimization problem. By adaptively reweighting penalties, IRFL directly targets sparsity in changepoint configurations, substantially improving support recovery and information criterion optimization compared to existing L1-based methods. Extensive simulations demonstrate that IRFL delivers reliable and accurate changepoint detection even in challenging settings with strong nuisance components, including trends, seasonality, and serial correlation. The framework naturally extends beyond time series to image data, enabling edge-preserving denoising and segmentation. Applications to real-world datasets highlight the practical impact of the approach. In particular, analysis of the Mauna Loa CO2 series reveals scientifically meaningful changepoints associated with volcanic eruptions and ENSO events, while providing a sharper and more interpretable trend decomposition than conventional methods.
