Change Detection for Streaming Data Using Wavelet-Based Least Squares Density–Difference (original) (raw)

2019

Abstract

Here, we present a novel algorithm for detecting changes in a continuous time series stream based on the \({\ell }_{2}\) distance between two distributions. The distributions are non-parametrically modeled using wavelet expansions, inspiring the name of our method: Wavelet-based Least Squares Density–Difference (WLSDD). Using the least squares method, we show that the \({\ell }_{2}\) distance between two wavelet expanded densities results in a closed-form expression of their coefficients. This circumvents the need to evaluate the densities and, instead, allows us to work directly with the differences between the corresponding scaling and wavelet coefficients. The method demonstrated superior change detection performance on both synthetic and real datasets, stationary or nonstationary, in comparison to other competing techniques.

O'Neil Smith hasn't uploaded this paper.

Let O'Neil know you want this paper to be uploaded.

Ask for this paper to be uploaded.