Clustering Time-series Data Based on the Modified Multiscale Matching Technique

摘要

This paper presents an improved version of time-series multiscale matching method that eludes the problem of shrinkage. The key idea is the development of new segment representation. The shape parameters of a Segment at high scale are now directly obtained using the shape parameters of base segments at the lowest scale, instead of using shapes represented by multiscale description. Multiscale shapes are now used only to obtain the hierarchy of the segments; since segment parameters are obtained independently of multiscale shapes, shrinkage does not distort them. We examined the usefulness of the method on the cylinder-bell-funnel dataset. The results demonstrated that the dissimilarity matrix produced by the proposed method, combined with conventional clustering techniques, lead to the successful clustering.