Our data set has $10^4$ data points, but has a long baseline and many gaps. As the histogram shows, the horizontal-axis is time and most of the time, there are no data. The vertical-axis is data counts. The total time is, in fact, short (green lines), but gaps make the time baseline very long.
enter image description here

If we bin the data, there would be $10^8$ data points[t,value], but only about $1\%$ are non-zero values. After binning, most of those values are zero because of those gaps.

How to improve detection efficiency(a faster method)?

A multi-threading way is possible(especially for Lomb-Scargle)?

  • $\begingroup$ Hi there, can you be a bit more specific about your dataset? How come you have more datapoints when binning them? What exactly do you mean by it having a long baseline? Do you have any plots or more information that we could work with to help you? $\endgroup$
    – MPIchael
    Apr 16 '19 at 9:21
  • $\begingroup$ Thanks for the attention. I revised the post and uploaded a test data set $\endgroup$ Apr 16 '19 at 14:32

One way to approach it, could be to do the analysis in two steps. First, you do a sweep along your dataset, and collect all nonzero values, and their times into a (much shorter!) data structure. You collect a list of tuples [t,value] basically. Every sweep after that will be extremely fast as you can safely assume every datapoint not in your list to be zero. Clicking on your link I have a hard time understanding what I see. These are not simple csv-style datapoints. They are in the form of:






Are these just your times and the second number is omitted?

How do you want to go about the periodicity detection? The obvious way here is to use a fourier transform which, after transform, will show you at which frequencies you have periodicity. Have you researched fast-fourier-transforms for python: (scipy.fftpack) ? I would be surprised if there would be no parallel implementation for a fourier transform in python. If that is still not fast enough there is the FFTW-library, boldly calling itself "Fastest Fourier Transform in The West", but I have to warn you that it is a pain in the ass to use. [EDIT:] There is a pyhton wrapper for FFTW pyFFTW.

You seem to have an impressive ammount of data. You might get away with doing some averaging, depending on your precision needs. If you take the first, say 10, datapoints, and take the maximum amplitude, store in a shorter array, and repeat for the following. You may shorten your dataset significantly (with loss in precision on your time axis of course).

  • $\begingroup$ They are just times. I need to bin those times(histogram) to get new Ts and values. fftpack is easy to get a significance level? $\endgroup$ Apr 18 '19 at 13:10
  • $\begingroup$ I am thoroughly confused. Does your data represent a set of timestamps of some event? Don't you need some amplitude or a second value to calculate a histogram? What are you trying to find out actually? $\endgroup$
    – MPIchael
    Apr 18 '19 at 13:18
  • $\begingroup$ They are time stamps. We bin the time stamps and get amplitudes. But that is not the point. The point is the time baseline is long and the data are gapped. $\endgroup$ Apr 19 '19 at 3:13

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.