Algorithm for clustering and visualization of non-normalized time-series data

Question

[sevenless]

I am looking for an algorithm, preferably implemented in R, that I can use to cluster non-normalized time-series data. My data set consists of scores that change over time which is why I want to keep the original values and avoid normalization. A very useful additional feature would be the visualization of the clustered time-series scores plotted together with their respective centroid.

I would be very grateful for any suggestions!

Answer 1

[Jean-Karim Heriche]

You can compute a matrix of distances between time series using dynamic time warping. Different variants of dynamic time warping are implemented in the R package dtw. You can then use the distance matrix with many clustering algorithms that accept a distance or similarity matrix as input. For visualization, you can use the distance matrix as input for multidimensional scaling (function cmdscale() or isoMDS() in the MASS package or layout.mds() in the igraph package).