VNU-UET Repository

Robust subspace tracking: Novel algorithm and performance guarantee

Le, Trung Thanh and Nguyen, Viet Dung and Nguyen, Linh Trung and Karim, Abed Meraim (2019) Robust subspace tracking: Novel algorithm and performance guarantee. Technical Report. VNU University of Engineering and Technology, Vietnam National University, Hanoi, Vietnam.

Full text not available from this repository. (Request a copy)

Abstract

Subspace tracking, which is refered to online PCA, is a classical problem in signal processing with various applications in wireless communications, rada and image/video processing. Since outliers and missing data are ubiquitous and more common in big data regime, robust variants of subspace tracking (RST) are crucial. In this paper, we propose a novel algorithm, namely PETRELSADMM, to improve RST performance in such scenario. The proposed approach consists of two main stages, including outlier rejection and subspace estimation. In the first stage, alternating direction method of multipliers (ADMM) solver is used to detect outliers residing in the observed data in an efficient way. In the second stage, we propose a modification of the parallel estimation and tracking by recursive least squares (PETRELS) algorithm to update the underlying subspace. A theoretical convegence analysis is provided, i.e., we prove that PETRELS-ADMM can generate a sequence of subspace solutions converging to the optimum of its batch counterpart. Performance studies show the superiority of our algorithms as compared to the state-of-the-art algorithms on both synthesis data and real data.

Item Type: Technical Report (Technical Report)
Additional Information: UET-AVITECH-2019003
Subjects: Electronics and Communications
Information Technology (IT)
Divisions: Advanced Insitute of Engineering and Technology (AVITECH)
Faculty of Electronics and Telecommunications (FET)
Depositing User: A/Prof. Linh Trung Nguyen
Date Deposited: 04 Dec 2019 07:30
Last Modified: 04 Dec 2019 07:30
URI: http://eprints.uet.vnu.edu.vn/eprints/id/eprint/3702

Actions (login required)

View Item View Item