A Robust Change-Point Detection Method by Eliminating Sparse Noises from Time Series

Singular Spectrum Transform (SST) is a fundamental subspace analysis technique which has been widely adopted for solving change-point detection (CPD) problems in information security applications. However, the performance of a SST based CPD algorithm is limited to the lack of robustness to corrupted observations with large noises in practice. Based on the observation that large noises in practical time series are generally sparse, in this paper, we study a combination of Robust Principal Component Analysis (RPCA) and SST to obtain a robust CPD algorithm dealing with sparse large noises. The sparse large noises are to be eliminated from observation trajectory matrices by performing a low-rank matrix recovery procedure of RPCA. The noise-eliminated matrices are then used to extract SST subspaces for CPD. The effectiveness of the proposed method is demonstrated through experiments based on both synthetic and real-world datasets. Experimental results show that the proposed method outperforms the competing state-of-the-arts in terms of detection accuracy for time series with sparse large noises.