Towards zero shot multivariate time series anomaly detection - A Realistic Evaluation

Ayan Banerjee, Sandeep Gupta

Keywords: multidimensional time series anomaly detection, zero shot, model recovery, model conformance

TL;DR: real world performance of anomaly detectors is evaluated on realistic case studies beyond standard benchmarks

Abstract: A long line of multivariate timeseries anomaly detection (MTAD) approaches use performance enhancement techniques that are not feasible in practical scenarios. In specific, a) point adjustment technique is employed which uses ground truth to forcefully convert false negatives to true positives and inflates precision to unrealistic proportions, and b) significant data leakage is introduced where anomaly score threshold is determined using the test data and test labels. In this paper, we show the real world performance of existing MTAD techniques when point adjustment and threshold learning on test data is disabled. Moreover, we show that anomalies introduced in real world benchmark datasets result in significant distribution shift between normal and anomalous data, and when point adjustment and threshold learning are used even untrained deterministic methods can perform on par or even beat baseline techniques. We then introduce six synthetic benchmark examples derived from real world systems, where anomalous data and normal data have statistically insignificant distribution shift. We propose, sparse model identification enhanced anomaly detection (SPIE-AD), a model recovery and conformance based zero shot MTAD approach that outperforms state of art MTAD techniques on three real world benchmark datasets without using point adjustment and threshold learning on test data. We evaluate state-of-art MTAD and SPIE-AD on the novel synthetic benchmarks. SPIE-AD outperforms state-of-art MTAD techniques on both standard and novel benchmarks.

Primary Area: learning on time series and dynamical systems

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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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.

Submission Number: 2727