Deep neural networks (DNNs) are attractive alternatives to more traditional methods for time series anomaly detection thanks to their capacity to automatically learn discriminative features. Despite their demonstrated power, different works have suggested that introducing engineered features in the time series can further improve the performance. In this work, we present a feature engineering strategy to transform univariate time series into a multivariate one by introducing non-local information in the augmented data. In this way, we aim to address an intrinsic limitation of the features learned by DNNs, which is they rely on local information only. We study the performance of our combination compared to each individual method and show that our method achieves better performance without increasing computational time on a set of 250 univariate time series proposed by the University of California, Riverside at the 2021 KDDCup competition.
From univariate to multivariate time series anomaly detection with non-local information
ECML/PKDD 2021, 6th Workshop on Advanced Analytics and Learning on Temporal Data, 13-17 September 2021 (Virtual Event) / Also to be published in LNCS, Vol. 13114
Poster / Demo
© Springer. Personal use of this material is permitted. The definitive version of this paper was published in ECML/PKDD 2021, 6th Workshop on Advanced Analytics and Learning on Temporal Data, 13-17 September 2021 (Virtual Event) / Also to be published in LNCS, Vol. 13114 and is available at : https://doi.org/10.1007/978-3-030-91445-5_12
PERMALINK : https://www.eurecom.fr/publication/6623