Mining Time Series for Identifying Unusual Sub-sequences with Applications

In a recent article, Eamonn et al. [2005] have introduced algorithms for the detection of most unusual time series sub-sequences. These have great implications for fast and intelligent data mining attempts using advances in modern computer technology. The techniques are used to detect unusual sub-sequences in time series arising from a wide range of applications. This paper is revisiting the algorithms introduced by the above authors and makes key improvements for a large class of time series processes by: (1) objectively identifying the size of the best sliding window for which similarities and discords could be found efficiently. (2) Reducing the processing time by a factor equivalent to the length of the best sliding window. (3) Introducing an entropy based measure as an alternative distance measure to account for outliers within specific sliding windows. (4) Highlighting comparisons with existing tools. (5) Demonstrating the new approach through applications on real life time series