COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection
Jinwoo Park, Hyeongwon Kang, Seung Hun Han, Pilsung Kang · 2026
Read COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection by Jinwoo Park, Hyeongwon Kang, Seung Hun Han, Pilsung Kang (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-2602-01635.txt.
Description
Jinwoo Park and colleagues build COMET for spotting anomalies in time series across settings, where single scale views and normal only training leave models blind to shifting distributions. COMET encodes patches at several scales to capture time dependencies and variable links, then learns normal patterns in a quantized coreset, scoring anomalies by quantization error plus memory distance. A third part adapts the codebook online at inference, generating pseudo labels and updating through contrastive learning. On five benchmark sets COMET tops thirty six of forty five metrics, showing strength across environments.
Catalog record
| CREATOR | Jinwoo Park, Hyeongwon Kang, Seung Hun Han, Pilsung Kang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 73033 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602.01635