Low Rank Transformer for Multivariate Time Series Anomaly Detection and Localization

Charalampos Shimillas, Kleanthis Malialis, Konstantinos Fokianos, Marios M. Polycarpou, [1ex] KIOS Research, Innovation Center of Excellence, Nicosia, Cyprus · 2026

Read Low Rank Transformer for Multivariate Time Series Anomaly Detection and Localization by Charalampos Shimillas, Kleanthis Malialis, Konstantinos Fokianos, Marios M. Polycarpou, [1ex] KIOS Research, Innovation Center of Excellence, Nicosia, Cyprus (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-08467.txt.

Description

Spotting faults in complex systems means finding both when something goes wrong and which variable is to blame. Charalampos Shimillas and three co-authors study how Transformers learn multi-variable time series, linking their attention to classical statistical methods. Their arXiv paper builds on those links to propose a low-rank attention model and a localization method.

The model, ALoRa-T, keeps self-attention patterns simple and low-rank, and its companion score captures the temporal shape of anomalies. ALoRa-Loc then attributes anomalies to specific variables by measuring links among series. Tests on simulated and real data show clear gains over leading methods on both detection and localization. Readers get theory-backed diagnostics for safety-critical monitoring.

Catalog record

CREATORCharalampos Shimillas, Kleanthis Malialis, Konstantinos Fokianos, Marios M. Polycarpou, [1ex] KIOS Research, Innovation Center of Excellence, Nicosia, Cyprus
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE89 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-08467

Browse the full library · About the Tower of Babel