Clustering short time series

Question

I would like to classify a relatively large set (over 9000) of short times series. The length of each sequence varies, but I would say about 80 % has between 2 and 9 observations. While I could use a simple trendline (maybe combined with a variance measure) to describe each these sequences, I would like to go a step beyond this solution.

What other kinds of methods could I utilize to cluster the "visual appearance" of these sequences? The ultimate goal of the classification is to gain a understanding of what type of trend/style/behavior each sequence is exhibiting.

The fact that the series differ in length greatly (greatly because they are short) will be a stumbling-block for any attempt of classification, even just visual one. — ttnphns, Mar 12 '12 at 15:28
Classify as in supervised learning a map to a given target class label? Or did you mean clustering? — Memming, Mar 12 '12 at 15:38
Did manage to use the wrong wording there: what I want to do is cluster. — Figaro, Mar 12 '12 at 16:45
Try running them through OPTICS with an appropriate dsitance function, e.g. DTW. — Has QUIT--Anony-Mousse, Mar 12 '12 at 17:28

score 1 · Answer 1 · edited Apr 13 '17 at 12:44

I can give you few hints:

You could use dynamic time warping to extract similarity between your sequences. Please see : Can someone please explain dynamic time warping for determining time series similarity?
Cave plot (visualization as you have stated) is another option. http://cm.bell-labs.com/cm/ms/departments/sia/project/misc/caveplot.html
If you use R, http://www.rdatamining.com/examples/ts-mining may give you some hint
You may want to see a discussion on similar topic at Time series 'clustering' in R

Clustering short time series

1 Answers1

Linked