Meaningful subsequence clustering for time series data stream
Vit Niennattrakul, Chotirat Ann Ratanamahatana · 2010
Subsequence clustering for time series data streams is one of the most challenging issues of time series data mining since subsequence clustering has been proven both theoretically and empirically that it produces meaningless clustering results, where hundreds of research works that utilize Subsequence Time Series Clustering (STSC) as a preprocessing step and a subroutine are all affected. Given a time series sequence, subsequence clustering should return cluster representatives which represent characteristics of all subsequences in time series. Therefore, if cluster representatives are always sine waves regardless of inputs, clustering results are meaningless since they do not reflect characteristics of the subsequences. The causes of meaninglessness are identified in twofold, i.e., inappropriate uses of Euclidean distance as a distance measure and Amplitude Averaging as an averaging function. To achieve meaningful clustering results, in this thesis, Shape-based Subsequence Time Series Clustering (2STSC) is proposed to use Dynamic Time Warping (DTW) distance measure and Shape-based Averaging function. Therefore, 2STSC returns more meaningful results than those from STSC. However, 2STSC cannot directly apply to data streams since 2STSC consumes large computational complexity by considering all previous subsequences for every new incoming data point. Shape-based Streaming Subsequence Time Series Clustering (3STSC) is then proposed to handle the streaming case by calculating a clustering result on a small set of stored subsequences instead of calculating from all previous subsequences. The small set of stored subsequences is updated for every new incoming data point to maintain the number of stored subsequences not to exceed the maximum allowance. 3STSC, therefore, is much faster than 2STSC, while 3STSC returns small distortions of clustering results.