Anomaly detection in time-series data has long been split into two camps: global methods, which judge a point against patterns spanning the whole dataset, and local methods, which focus on behavior within a narrow neighborhood. Researchers Tongkai Lu, Shuai Ma, Zhongxi Zhang, Lizhen Cui and Xuelian Lin report in Knowledge and Information Systems that, to their knowledge, no prior study has systematically combined the two approaches for time-series data in a single framework. Their new method does exactly that, and the reported benchmark results suggest the combination pays off in both accuracy and speed.
The unification rests on a discretization step. The authors convert continuous time-series data into discrete form and then partition the discrete data between a global anomaly-detection component and a local one, using a concept they call the local factor to decide which points each part handles. This division allows each detector to work on the portion of the data it is best suited to, rather than forcing one method to cover every kind of anomaly.
On the global side, the team designed a rule-based method built on non-redundant association rules with tolerances. According to the abstract, this design targets two specific difficulties in time-series data: the latency phenomenon, where anomalous effects appear with a delay, and the excessive number of rules that naive rule-mining can produce. Restricting the rule set to non-redundant rules with tolerances is intended to keep the global detector both tractable and sensitive to delayed effects.
The local component takes a different route. It is a Transformer-based anomaly detector that incorporates LSH (Locality Sensitive Hashing) Attention, a mechanism the authors say fits discrete data better than standard attention.
The empirical claims are concrete. The same comparisons showed running-time reductions of 43.0%, 47.0% and 60.3% respectively, meaning the accuracy gains came alongside substantially faster execution rather than a speed-accuracy trade-off.
The evaluation extended beyond those three datasets. Across 48 datasets from the TimeEval benchmark suite, the new method reached a mean F1 score of 89.19%, against 84.34% for Anomaly Transformer. The authors describe the experimental study as extensive and say it was conducted to verify both the effectiveness and the efficiency of their approach.
The reported numbers come from the article’s abstract, which summarizes the study published online on 28 October 2026 as Volume 68, article number 299 of the journal. Full methodological details sit behind the article’s subscription access, so independent replication details, such as hyperparameter choices and statistical tests, cannot be verified from the preview alone.
Subject of Research: A unified framework combining global and local anomaly detection methods for time-series data.
Article Title: Unifying global and local anomaly detection for time series
Article References: Lu, T., Ma, S., Zhang, Z., Cui, L., & Lin, X. (2026). Unifying global and local anomaly detection for time series. Knowledge and Information Systems, 68(1), Article 299. https://doi.org/10.1007/s10115-026-02917-9
Image Credits: AI Generated
DOI: 10.1007/s10115-026-02917-9
Keywords: anomaly detection, time series, machine learning, Transformer, association rules, LSH attention, benchmark evaluation, industrial security, Knowledge and Information Systems, Unifying, global, local
News Source: Denise Maddox. (October 29, 2026). One framework joins global and local anomaly detection for time series. Scienmag.
