0
$\begingroup$

I have univariate time series data for 70 subjects sampled at 1000 Hz. When graphing the single subject plots, time is on the x-axis and amplitude (arbitarty unit) on the y-axis. When looking at the data, I'm seeing 3 "types" of subjects but there is no hard way of classifying the subjects. I want to use an unsupervised machine learning classification approach to put all of these subjects in 3 "classes" or "types" of subjects. I was thinking about using dynamic time warping or recurrent neural networks, but am not sure if this is the best approach. Could someone please help me with this? Maybe provide me with some info on how to construct these models/perform these analyses?

Thanks in advnace!

$\endgroup$
1
$\begingroup$

Since you are dealing with a small number of time series, I think recurrent neural networks are a bit over the top. I would start with a simple approach and check if the results match your expectations, if not refine it. As a start clustering could be the way to go. Since you are expecting 3 "types" of time series, you could use K-Medoid Clustering setting K to 3. As a distance metric you already mentioned DTW, but depending on your definition of similarity in this context, other distance measures might be more suitable.

I know that the ELKI Data Mining Framework has both components already implemented. Also you can quickly change distance functions if you want to.

$\endgroup$
1
  • $\begingroup$ @Ravi If you feel that my answer helped you, you could accept my answer :) $\endgroup$
    – SaiBot
    Dec 19 '17 at 7:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.