130 likes | 397 Views
Novel Online Methods for Time Series Segmentation. Xiaoyan Liu, Zhenjiang Lin, and Huaiqing Wang TKDE, Vol. 20, No. 12, 2008, pp. 1616-1626. Presenter : Wei-Shen Tai 200 9 / 1/20. Outline . Introduction Related work Novel online segmentation algorithms: FSW & SFSW Complexity analysis
E N D
Novel Online Methods for Time Series Segmentation Xiaoyan Liu, Zhenjiang Lin, and Huaiqing Wang TKDE, Vol. 20, No. 12, 2008, pp. 1616-1626. Presenter : Wei-Shen Tai 2009/1/20
Outline • Introduction • Related work • Novel online segmentation algorithms: FSW & SFSW • Complexity analysis • Experiments • Conclusions • Comments
Motivation • Represent a time series approximately in a few segments • Representation quality : minimizing the representation error as possible. • Computing efficiency : fast enough to fit for an online real-time working environment.
Objective • Novel online segmentation algorithms • Efficiently finds the farthest endpoint of a segment and reduces the representation error for dealing with an online data sequence.
Sliding window method • Classic SW • Interpolating line or regression line between the two endpoints of the segment is used as the approximation. t1 • t2 • t3 • t4
Segmentation criterion • Evaluation for the goodness of fit line in segmentation methods • Regression line • Residual error • Interpolation line • MVD: sum of the squares of vertical distances between actual data points and the best fit line. • Maximum error tolerance • A user-specified maximum error tolerance δ.
Feasible Space Window (FSW) • Candidate Segmenting Point (CSP) • Chosen to be the next eligible segmenting point. • FSW • Searches for the farthest CSP to make the current segment as long as possible under the given maximum error tolerance.
Stepwise Feasible Space Window (SFSW) • SW method • Lacks an overall view of the whole time series. • SFSW • Backward FSW to find a backward segmenting end point. • Find the optimal segmenting point in the interval of both forward and backward end points.
Complexity analysis • Comparison between SW methods • Given time series T of n data points, the number of segments and the average segment length are denoted by K and L, respectively.
Conclusions • FSW • Reduces the number of segments by searching for the farthest endpoint of a potential segment. • SFSW • Refines the segmenting points by taking into account the effects of new incoming points so that the representation error can be reduced. • Future works • An amnesic representation, varying stepwise method, continuous feature discreteziation, multidimensional time series.
Comments • Advantage • This method superiors to other SW methods in that it can consider the more global view of time series via backward FSW. • The user-specified threshold δ and feasible space concept make this method become an incremental segmentation algorithm. • Drawback • FSW is an quite efficient method but its representation error is larger than other SW methods. • SFSW can reduce both the number of segment and representation error but increase its computation complexity also. • Application • Time series segmentation.