The best is to have real data to analyse but it can be very useful to simulate these pattern in order to assess the clustering efficiency.Ī simple way to simulate these time series is to use the sine function and to add a random noise in order to make it more credible. “hitting the wall”) with a jump in the pace which indicates a significant slow down in the second part of the run. Let’s imagine two people running a marathon, one had a classic run with a pace increasing with the time and the other had a very bad experience (e.g. To make this process clearer I’m going to simulate two groups of time series and to check if whether or not the DTW clustering can differentiate them. However when it comes to analyse real data, I found difficult to understand how the clustering is working. These packages are very simple but powerful tools to analyse time series. The main advantage of DTW is the possibility to group time series according their patterns or shapes even if these patterns are not synchronized (lag).Īs far as I know the two main packages which allow time series clustering with DTW are TSclust by Pablo Montero Manso and José Antonio Vilar and dtwclust by Alexis Sarda-Espinosa. Collier’s blog post, Peter Laurinec’s blog post, Dylan Glotzer’s lecture or Ana Rita Marques’s module.ĭynamic Time Warping (DTW) is one of these solutions. Many solutions for clustering time series are available with R and as usual the web is full of nice tutorials like Thomas Girke’s blog post, Rafael Irizarry and Michael Love’s book, Andrew B.

Time series clustering