EcoSta 2024: Start Registration
View Submission - EcoSta2024
A0844
Title: Robust singular value decomposition Authors:  Ayanendranath Basu - Indian Statistical Institute (India) [presenting]
Subhrajyoty Roy - Indian Statistical Institute (India)
Abhik Ghosh - Indian Statistical Institute (India)
Abstract: Singular value decomposition (SVD) of a data matrix is traditionally based on the least squares principle and, as a consequence, is very sensitive to the presence of outliers. As a result, the different application domains that use classical methods of SVD may experience degraded performance. A robust singular value decomposition technique is proposed based on the minimum density power divergence estimator. Apart from the theoretical properties of the estimator, a practical algorithm based on alternating weighted regression is also proposed to obtain the estimate. Simulation results and a real application of the video surveillance background modelling problem are presented to demonstrate the performance of the method.