Few-shot time-series anomaly detection via temporal clustering

Abstract

Time-series anomaly detection is a study that finds unusual data in a series of observed, chronologically listed data. Anomaly detection in time-series data has become an important task in many applications. It can provide valuable insights into the underlying process or system being monitored, helping to prevent catastrophic events, reduce costs, and improve overall efficiency. As time-series data becomes increasingly complex and larger, there are some limitations in the quality of time-series data, such as a lack of labels, inaccurate labels, and contamination of training data. Therefore, in this work, we propose to cluster dimensions of multivariate time series into meaningful groups to train anomaly detection models for each cluster. We also take advantage of a few anomaly labels included in the training dataset to proceed with this process. We confirm that real-world data can be separated into several groups that share similar patterns and that, in fact, anomaly detection performance increases when the model is trained for each cluster.