Abstract:
|
Even though statistical clustering algorithms go back to the 1950s, development has continued unabated with particular interest from the statistical learning community. But time series has been a blind spot with only a small statistical literature. And while the statistical learning literature apparently treats 'time series' this is not done by modelling autocorrelation but by 'warping' methods. Here we exhibit a novel clustering algorithm obtained, like k-means, as a limiting form of a mixture of autoregressions clustering method. While it shows similar performance to the mixture classifier it is orders of magnitude faster.
|