Linear-trend normalization for multivariate subsequence similarity search
Thibaut Germain, Charles Truong, Laurent Oudre · 2024
Finding repeating or anomalous subsequences in long time series is a crucial task in numerous data analysis pipelines. Most of those methods share a common step where they compute the pairwise similarity between all subsequences of a time series or between a fixed subsequence and a time series. However, the presence of a trend in a time series may cause changes in the shape of subsequences, making the similarity measure less reliable. This article introduces a new normalization scheme called LT-normalization (for Linear Trend) to prevent this phenomenon. It generalizes the well-known Z-normalization by removing the linear trend and scaling the subsequences to unit variance. Like the Z-normalization, we show that the LTnormalization has a computationally efficient recursive formulation. Thanks to this recursion property, the LT-normalized matrix profile can be computed with the same quadratic complexity as the classical Z-normalized matrix profile. Our procedure can naturally cope with multivariate signals. Empirical results on synthetic and real datasets show that the LT-normalized matrix profile has competitive performances for the best motif pair, similarity search, and motif set discovery problems.