Vector quantization: A discretization technique for fast time series discord discovery

Time series discords are defined as subsequences of a longer time series that are maximally different to all of the rest of the time series subsequences under a predefined distance function. Finding time series discords plays a very important role in detecting anomalies in data gathered from various application domains, such as space shuttle telemetry, mechanical industry, as well as biomedicine. Although there have been several algorithms for finding time series discords proposed in the literature, they still suffer from very high runtime. In many applications, the performance bottleneck is truly undesired since any unusual event should be discovered as fast as possible. In order to tackle this problem, we propose a new approach for speeding up the discord discovery process which is built upon the vector quantization technique to reduce dimensions and discretize time series data. Experimental results show that our proposed approach outperforms HOT SAX and WAT, the two state-of-the-art algorithms were also introduced in the literature recently.