I have over 1000 samples with each sample itself being a multivariate time series data (of rows between 500 and few thousands). The length of time series differs from sample to sample. What are the possible approaches to perform an unsupervised clustering such that each cluster can describe "Average time series behavior" for all samples in the same cluster?
You can normalise all the samples to be with the same length, then you can use any clustering technique such as k-means, hierarchical clustering, density-based clustering or subspace clustering, to find clustering structures.
Then use can work on each cluster/groups by combining, averaging them..etc.
Eiman Kanjo Thanks for answering. The time series data I have can be normalized for sure but before applying any clustering technique, one would need to decompress these vertical time series into a set of features horizontally. Some techniques I had in mind are things like FFT based features or even simple descriptive statistics like std, kurtosis, mean etc. And this is where I am stuck in. What should be the right approach of generating these features.
I have done univariate time series clustering in SAS. See SAS Documentation or paper by Goutam Chakraborty et al. http://support.sas.com/resources/papers/proceedings12/200-2012.pdf
I found the following article which introduces " a novel approach based on dynamic time wrapping and parametric Minkowski model has been proposed"
B. Chandra, M. Gupta and M. P. Gupta, "A multivariate time series clustering approach for crime trends prediction," 2008 IEEE International Conference on Systems, Man and Cybernetics, Singapore, 2008, pp. 892-896. doi: 10.1109/ICSMC.2008.4811393 URL: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=4811393&isnumber=4811240
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts