Some efficient segmentation-based techniques to improve time series discord discovery

Abstract

Time series discord has proved to be a useful concept for time series anomaly detection. To search for discords, various algorithms have been developed. HOT SAX has been considered as a well-known and effective algorithm in time series discord discovery. However this algorithm still has some weaknesses. First, users of HOT SAX are required to choose suitable values for the discord length, word-length and/or alphabet-size, which are unknown. Second, HOT SAX still suffers from high computation cost. In this paper, we propose some novel techniques to improve HOT SAX algorithm. These techniques consist of (i) using some time series segmentation methods to estimate the two important parameters: discord length and word length and (ii) speeding up the discord discovery process by a new way of shifting the sliding window. Extensive experiments have demonstrated that the proposed approach can not only facilitate users in setting the parameters, but also improve the discord discovery in terms of accuracy and computational efficiency.