Window Projection Features are All You Need for Time Series Anomaly Detection

Issar Tzachor, Yedid Hoshen

Keywords: time series, anomaly detection

TL;DR: A simple hand-crafted representation combined with a Gaussian estimator obtains SOTA results in time series anomaly detection.

Abstract: The challenge of time series anomaly detection has motivated the development of increasingly more complex deep representations and anomaly metrics. In this paper we demonstrate that a simple approach based on window projection features can achieve better results. Projection features are a common way to discretize multivariate data; they first multiply the data by a projection matrix followed by discretization of each output dimension. We first show that short temporal windows, encoded by projection features, are often already sufficiently expressive for linearly separating between normal and anomalous time series. However, we find that while the expressivity of projection features is sufficient, current one-class classification methods are unable to use them effectively to detect anomalies. We hypothesize this is due to the difficulty of density estimation. The difficulty can be overcome by estimating the probability density of the sample mean, which follows the Gaussian distribution when the conditions of the Central Limit Theorem are met. Simply put, we fit a multivariate Gaussian model to the average of the projection features of adjacent windows within a time series. Despite its simplicity, our method outperforms the state-of-the-art in diverse settings including: five UEA datasets, video trajectory anomaly detection and standard anomaly segmentation benchmarks. Code is provided.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Unsupervised and Self-supervised learning