Anomaly Detection in Univariate Time Series: HOT SAX vesus LSTM-Based Method

Anomaly detection in time series has been an important and challenging research topic. There have been several methods proposed for detecting anomaly subsequences in a time series. The majority of these methods is classified into the window-based category, which applies a sliding window with a fixed length to extract subsequences before finding out anomaly subsequences. A well-known algorithm in this category is HOT SAX algorithm. Recently, deep neural network models, especially Long Short Term Memory (LSTM) network, are also applied for time series anomaly discovery. LSTM-based methods for time series anomaly detection belong to prediction-based category. So far, there has been no research work to compare the performance of LSTM-based method to that of any traditional window-based method in time series anomaly detection. The research question investigated in this paper is that whether the newly developed LSTM-based method for time series anomaly detection is superior to the traditional algorithms, such as HOT SAX or not. In this study, we give an empirical comparison between LSTM-based method and HOT SAX in time series anomaly detection. Extensive experiments on seven benchmark time series indicate that LSTM-based method is not superior to HOT SAX since it brings out the same detection accuracy as HOT SAX while it incurs much higher computational overhead.