0
$\begingroup$

I need to find a specific pattern in a continual time series and then to classify it into one of two groups good, or bad. So, the first step would be some kind of search in the time series, but that also comes to classification.

My question is not related to the classification per se, but to 'preparation' to classify. Concretely, as this is a stream of data my first question would be how to determine the size of the window? And whether these windows should overlap and by how much? I'm afraid that the result will pretty much depend on the way I perform the windowing.

Another thing which bugs me is how to prepare the data,as it is not labeled. Would it make sence to split the series, cluster it and then learn on specific cluster. The other approach I have in mind is to create a couple of samples manually, by splitting on windows which represent 'nice' examples and then perform some kind of bootstraping where I classify the stream,identify pattern and add that window to the training set. Which approach would be better? And is there any other way?

Improve this question
$\endgroup$
4
  • $\begingroup$ Is the specific pattern which you are trying to find predefined ? Do you know what you are looking for? Or are you looking for any motifs or novelties? $\endgroup$
    – Nikolas Rieble
    Commented Nov 12, 2016 at 8:18
  • $\begingroup$ I know what I'mblooking for, two log shaped patterns, which can last longer or shorter and get to greater or smaller values, but they are log shaped. I don't have labeled samples, just a sample stream and an explanation what is of interest to isolate. Later there will be some novelty detection, most probably, but first this. $\endgroup$
    – Kobe-Wan Kenobi
    Commented Nov 12, 2016 at 8:36
  • $\begingroup$ Deal with both steps seperately. First is pattern matching / extraction. Second is classification. Once you have set up both steps, you can work on the effect of the window size onto the classification performance $\endgroup$
    – Nikolas Rieble
    Commented Nov 12, 2016 at 9:22
  • $\begingroup$ Generally the window size must most certainly be longer than the target pattern. Further, the windows must overlap. What is the sampling rate of your incoming stream? $\endgroup$
    – Nikolas Rieble
    Commented Nov 12, 2016 at 9:24

0

Reset to default

Your Answer

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