Empirical Comparison of Clustering Methods for Long Time-Series Databases

In this paper we report some characteristics of time-series comparison methods and clustering methods found empirically using a real-world medical database. First, we examined basic characteristics of two sequence comparison methods, multiscale matching (MSM) and dynamic time warping (DTW), using a simple sine wave and its variants. Next, we examined the characteristics of various combinations of sequence comparison methods and clustering methods, in terms of interpretability of generating clusters, using a time-series medical database. Although the subjects for comparison were limited, the results demonstrated that (1) shape representation parameters in MSM could capture the structural feature of time series; for example, the difference of amplitude was successfully captured using rotation term, and that differences on phases and trends were also successfully reflected in the dissimilarity. (2) However, the dissimilarity induced by MSM lacks linearity compared with DTW. It was also demonstrated that (1) complete-linkage criterion (CL-AHC) outperforms average-linkage (AL-AHC) criterion in terms of the interpret-ability of a dendrogram and clustering results, (2) combination of DTW and CL-AHC constantly produced interpretable results, (3) combination of DTW and RC would be used to find core sequences of the clusters. MSM may suffer from the problem of 'no-match' pairs, however, the problem may be eluded by using RC as a subsequent grouping method.