Comparison of similarity measures and clustering methods for time-series medical data mining

This paper reports characteristics of dissimilarity measures used in the multiscale matching. Multiscale matching is a method for comparing two planar curves by partially changing observation scales. Throughout all scales, it finds the best set of pairs of partial contours that contains no miss-matched or over-matched contours and that minimizes the accumulated differences between the partial contours. In order to make this method applicable to comparison of the temporal sequences, we have proposed a dissimilarity measure that compares subsequences according to the following aspects: rotation angle, length, phase and gradient. However, it empirically became apparent that it was difficult to understand from the results that which aspects were really contributed to the resultant dissimilarity of the sequences. In order to investigate fundamental characteristics of the dissimilarity measure, we performed quantitative analysis of the induced dissimilarities using simple sine wave and its variants. The results showed that differences on the amplitude, phase and trends were respectively captured by the terms on rotation angle, phase and gradient, although they also showed weakness on the linearity.