A Joint Soft Warping and Clustering Approach to Detecting Time Series Anomalies

Abstract

Unsupervised anomalous time series detection methods focus on identifying outliers without prior knowledge of the dataset. However, these methods often require multiple parameters to be optimized, with adequate performance tied to their careful tuning and prior domain knowledge. In this work, two methods are proposed for detecting outlier time series that adopt a joint clustering and alignment optimization to filter out the desired signals. The time series are globally clustered while simultaneously being aligned to other signals in their same cluster group. This alternating optimization employs time-warping similarity measures to help identify closely matching time series as well as the outliers. The proposed techniques require minimal parameter tuning and yield superior results on many benchmark datasets.