1
$\begingroup$

I have different univariate time series and the goal is to detect outliers automatically. Therefore I used different algorithms for different time series. But the first step would be to detect automatically the time series class. Attached you could see different univariate time series (continuous1, binary, continuous2).

HOW could I distinguish between these?

I have already tried kmeans with some simple generated features but it doesn't work. The first main goal should be to distinguish between binary and continuous data.

enter image description here

enter image description here

enter image description here

$\endgroup$
2
  • $\begingroup$ I think it depends a lot on your outlier detection algos. If the algos are informed by the scientific knowledge of the DGPs then ideally the same knowledge should be used to automatically categorise time series into different DGP. OTOH, if this is purely statistical exercise, variance / autocorrelations / range or other such measures can be explored. Bottomline is that you want the clustering to be such that each cluster of time series represent a particular class of DGP for which the concerned outlier detection algo with work better as compared to algo applied for other clusters. $\endgroup$
    – Dayne
    Commented Dec 1, 2021 at 7:41
  • $\begingroup$ For the immediate requirement of binary and continuous, you can just put a check for number of unique values and segregate. $\endgroup$
    – Dayne
    Commented Dec 1, 2021 at 7:44

1 Answer 1

0
$\begingroup$

First of all clustering of time-series with usual clustering algorithm tends to give improber results, which is demonstrated quite clearly in this paper. However, you could aim for clustering describing parts of time-series; like seasonal frequency, trend, cycle and so on. This time-series specific attributes themselves can be clustered and therefore may help you in finding "similiar" ts. A R-Package with a good time-series clustering function is timetk.

$\endgroup$
1
  • $\begingroup$ the paper you have cited is about clustering within a timeseries (subsequences). The OP is about clustering a group of timeseries into different clusters so that appropriate outlier detection algo can be used. $\endgroup$
    – Dayne
    Commented Dec 1, 2021 at 8:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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