Memory-efficient centroid decomposition for long time series

Real world applications that deal with time series data often rely on matrix decomposition techniques, such as the Singular Value Decomposition (SVD). The Centroid Decomposition (CD) approximates the Singular Value Decomposition, but does not scale to long time series because of the quadratic space complexity of the sign vector computation. In this paper, we propose a greedy algorithm, termed Scalable Sign Vector (SSV), to efficiently determine sign vectors for CD applications with long time series, i.e., where the number of rows (observations) is much larger than the number of columns (time series). The SSV algorithm starts with a sign vector consisting of only 1s and iteratively changes the sign of the element that maximizes the benefit. The space complexity of the SSV algorithm is linear in the length of the time series. We provide proofs for the scalability, the termination and the correctness of the SSV algorithm. Experiments with real world hydrological time series and data sets from the UCR repository validate the analytical results and show the scalability of SSV.