A1429
Title: A semi-parametric approach for clustering high-dimensional, non-stationary, auto-correlated time series
Authors: Qiyuan Wang - Texas A&M University (United States) [presenting]
Abstract: A novel semi-parametric estimation algorithm is developed for accurately estimating time-varying mean and variance in autoregressive (AR) models. Utilizing B-splines with generalized least squares (GLS) estimation for smooth parametrization and weighted least squares (WLS) for more precise estimation, the approach addresses the challenges posed by time-varying dynamics in time series data. The covariance matrix in the GLS estimation of the spline coefficients is iteratively updated by calculating it through the WLS estimation of the AR coefficients in a band-limited manner. Meanwhile, a new autoregressive model is proposed that incorporates time-varying variance with a finite bounded envelope function, and a novel method is introduced to estimate it through splines. Additionally, the order of the AR model is determined through a generalized Bayesian information criterion (GBICp) that incorporates prior information. The effectiveness of the methodology is demonstrated through extensive simulations and applications to real-world electrocardiograms (ECGs) data, showcasing significant improvements in the dimension reduction while preserving major features for high-accuracy clustering tasks.
