4
$\begingroup$

I'm looking for some good pointers to pattern recognition with time series. Possibly something practical that can be easily understood.
As a toy example, think about collecting data from an accelerometer of a phone. Different types of motion (running, sitting, falling down, etc..) may correspond to different patterns. This is just an example, and I'm looking for some feasible solutions in the domains of time series.

Of course, if there are methods not based on time series, I'm open also.

$\endgroup$
  • 1
    $\begingroup$ For a worked example with data akin to yours please visit mathematica.stackexchange.com/questions/2711/…. The question stands out as particularly broad because "pattern" could mean practically anything. Even a full answer to "how do I detect a sudden change in mean" would take some effort: that's the subject of change-point analysis. $\endgroup$ – whuber Oct 7 '14 at 21:17
  • $\begingroup$ Bob, when you mean "pattern" are you referring to unsupervised learning/clustering or supervised learning ? $\endgroup$ – forecaster Oct 7 '14 at 23:07
  • $\begingroup$ possibly unsupervised, but i'm open to supervised $\endgroup$ – Bob Oct 8 '14 at 1:59
1
$\begingroup$

Pattern recognition in time series can involve a number of components. Memory i.e. auto-dependence can be characterized via an ARIMA component (stochastic/adaptive structure). Deterministic structure such as level shifts,local time trends,pulses and seasonal pulses can be found via Intervention Detection schemes. Changes in error variance and changes in parameters of the model over time can be found via residual diagnostic checking. Furthermore patterns can respond to known events , oftentimes the response is in anticipation of the event but much more frequently when the event occurs and periods following the event. If you have a pet time series please post it and I will try and develop the underlying pattern via available software.

$\endgroup$
  • $\begingroup$ Thanks a lot for the information. I cannot post the data I use. However, I will google the keywords you mentioned (I'm not an expert in time series) and will try to apply this stuff. $\endgroup$ – Bob Oct 8 '14 at 2:02
  • $\begingroup$ If I can help further i.e. more specific references I would be glad to. You can learn a lot about what I have said by reading my prior posts as I only comment about time series methods. $\endgroup$ – IrishStat Oct 8 '14 at 12:25
  • $\begingroup$ It'd be great if you could add some references in the literature. I need some good tutorial to get a grasp in this matter. Thank you! $\endgroup$ – Bob Oct 8 '14 at 18:14
  • 1
    $\begingroup$ ARIMA MODEL Identification; people.duke.edu/~rnau/arimrule.htm Intervention Detection and error variance change detection unc.edu/~jbhill/tsay.pdf Parameter Change Detection en.wikipedia.org/wiki/Chow_test $\endgroup$ – IrishStat Oct 8 '14 at 20:10
  • 1
    $\begingroup$ There is a useful tutorial here autobox.com/cms/index.php/afs-university/intro-to-forecasting $\endgroup$ – IrishStat Oct 11 '14 at 16:15
0
$\begingroup$

correlogram (corrgram package in R), harmonic analysis, spectral analysis (eg spec.pgram in R), depending on how you want to decompose the pattern(s). auto.arima {forecast} is a useful implementation based on the suggestion above. assuming you have univariate response.

$\endgroup$

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.