Subscribe Bookmark RSS Feed

Experiences clustering time series data using JMP?

gv

Community Trekker

Joined:

Jul 28, 2014

Both for simulation and real world data it can be of great value to identify time series that are similar in both their absolute values over time but also in terms of their modes of behaviour. For example, this can help us identify when the behaviour of a system is drastically changed in response to a change in input. JMP offers some very useful clustering techniques but as far as I am aware none specifically for time-series data.

Has anyone ever attempted to develop a script or plugin to cluster time series data in JMP? I know such packages exist for R. For example, TSclust.

A very crude approach is feeding the JMP clustering algorithm time series data with each data point in a new column (so column 1 = T1 .. etc.). This is a quick and dirty approach that already delivers some results. But I am looking for an accessible way to more accurately identify similarities in behaviour. For example: a graph could be divided in segments and for each segment we could identify if the behaviour shows a certain types of behaviour: accelerating growth, accelerating decline (exponential) or decelerating growth and decelerating decline (logarithmic). Clustering of this information could then be done by the normal clustering algorithms included in JMP. However, ideally the clustering is done for the whole sequence of behaviour (which might include many changes in the shape of the behaviour) while also looking at various features of the behaviour, for example, the first and second derivative as well as the average value. I am interested to know if anyone has experiences with this challenge and is willing to share.

Attached you can find a very crude approach using Ward clustering. The results are actually not to bad (eyeball test) (See image attached). The data table with the clustering script included is attached:

11308_pastedImage_2.png

1 REPLY
billw_jmp

Staff

Joined:

Jul 2, 2014

gv,

I stacked your data and then did a control chart with phasing for the clusters.  Not sure if this is helpful or useful, but thought you might like it as an alternative.

11315_pastedImage_0.png