1
$\begingroup$

Let's say I have discrete events in time, e.g. patients getting sick, and I want to predict whether theses events are indicators of some other underlying event, e.g. a disease outbreak. Usually, one would transform the event based time series into a regular time series by for example aggregating the counts of patients over each week. However, I feel that by this aggregation a lot of information is lost and it is also hard use to multiple features of events, e.g. patients' age, patients' sex, etc. My question is very general: is there a branch of statistics / data science that treats time series as composed of events at arbitrary times rather that aggregated values at evenly distributed intervals?

I tried googling it, but it seemed hard to phrase the question in a way that a search engine understands.

Improve this question
$\endgroup$
1
  • 1
    $\begingroup$ Nonstationary hidden Markov model perhaps? $\endgroup$
    – HEITZ
    Nov 14, 2018 at 21:15

2 Answers 2

Reset to default
2
$\begingroup$

"rather than aggregated values at evenly distributed intervals?" I think that part defines your question very well; i.e that is usually what we generally do in usual time-series problems, merging the events in some discrete time intervals.

What you can do is that, you can give the time distance between the events as features, without giving evenly distributed intervals. For example, the exact time distances between events of every disease outbreak or every patient getting sick and etc. Moreover, you can do some feature engineering; squared time differences or percentage changes of time difference between the differences between the sequential doubles (x, x-1) and (x-1, x-2) where x-n,.....,x-2, x-1, x are the sequential events, not the time.

The relevant concept here seems to be "unevenly spaced time-series", the wiki page Wikipedia: Unevenly spaced time series and this researcher's website http://eckner.com/research.html seems to give enough material to get started with.

Improve this answer
$\endgroup$
4
  • 2
    $\begingroup$ Yes using the times or time distances as features was exactly what I was looking for! Non-periodic time series was not really right, but from there I got to "unevenly spaced time-series" which seems to be the correct term. The wiki page en.wikipedia.org/wiki/Unevenly_spaced_time_series and this guys page eckner.com/research.html seem to give enough material to get started with. Thanks! If you incorporate that into your answer, I can accept it. $\endgroup$
    – Jarno
    Nov 16, 2018 at 11:06
  • $\begingroup$ Just done. Glad I could help, at least with some part. Good luck! $\endgroup$
    – Ugur MULUK
    Nov 16, 2018 at 11:29
  • 1
    $\begingroup$ Also called "irregular time series" and for which the its R package was developed (but appears to have been abandoned). I see that the cited work by Eckner also uses R. $\endgroup$
    – 42-
    Nov 20, 2018 at 18:15
  • $\begingroup$ Despite having been moved to the CRAN Archive, the its package still installs with install.packages("~/Downloads/its_1.1.8.tar.gz", repo=NULL) and loads without complaints. $\endgroup$
    – 42-
    Nov 20, 2018 at 18:25
1
$\begingroup$

That looks like Duration Modelling, a.k.a. Survival Analysis. You can find an R tutorial here.

Improve this answer
$\endgroup$

Your Answer

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

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