TY - GEN
T1 - TS-Bert
T2 - 21st International Conference on Computational Science, ICCS 2021
AU - Dang, Weixia
AU - Zhou, Biyu
AU - Wei, Lingwei
AU - Zhang, Weigang
AU - Yang, Ziang
AU - Hu, Songlin
N1 - Publisher Copyright:
© 2021, Springer Nature Switzerland AG.
PY - 2021
Y1 - 2021
N2 - Anomaly detection of time series is of great importance in data mining research. Current state of the art suffer from scalability, over reliance on labels and high false positives. To this end, a novel framework, named TS-Bert, is proposed in this paper. TS-Bert is based on pre-training model Bert and consists of two phases, accordingly. In the pre-training phase, the model learns the behavior features of the time series from massive unlabeled data. In the fine-tuning phase, the model is fine-tuned based on the target dataset. Since the Bert model is not designed for the time series anomaly detection task, we have made some modifications thus to improve the detection accuracy. Furthermore, we have removed the dependency of the model on labeled data so that TS-Bert is unsupervised. Experiments on the public data set KPI and yahoo demonstrate that TS-Bert has significantly improved the f1 value compared to the current state-of-the-art unsupervised learning models.
AB - Anomaly detection of time series is of great importance in data mining research. Current state of the art suffer from scalability, over reliance on labels and high false positives. To this end, a novel framework, named TS-Bert, is proposed in this paper. TS-Bert is based on pre-training model Bert and consists of two phases, accordingly. In the pre-training phase, the model learns the behavior features of the time series from massive unlabeled data. In the fine-tuning phase, the model is fine-tuned based on the target dataset. Since the Bert model is not designed for the time series anomaly detection task, we have made some modifications thus to improve the detection accuracy. Furthermore, we have removed the dependency of the model on labeled data so that TS-Bert is unsupervised. Experiments on the public data set KPI and yahoo demonstrate that TS-Bert has significantly improved the f1 value compared to the current state-of-the-art unsupervised learning models.
KW - Anomaly detection
KW - Pre-training model
KW - Time series analysis
UR - https://www.scopus.com/pages/publications/85111399947
U2 - 10.1007/978-3-030-77964-1_17
DO - 10.1007/978-3-030-77964-1_17
M3 - 会议稿件
AN - SCOPUS:85111399947
SN - 9783030779634
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 209
EP - 223
BT - Computational Science – ICCS 2021 - 21st International Conference, Proceedings
A2 - Paszynski, Maciej
A2 - Kranzlmüller, Dieter
A2 - Kranzlmüller, Dieter
A2 - Krzhizhanovskaya, Valeria V.
A2 - Dongarra, Jack J.
A2 - Sloot, Peter M.A.
A2 - Sloot, Peter M.A.
A2 - Sloot, Peter M.A.
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 16 June 2021 through 18 June 2021
ER -