Suppose, I have a 30 seconds time-step observations of sports data, in some of the intervals the game was partially/fully stopped. I'm trying to prep the data for a time series analysis. Is it justified to take it as zero when it was stopped fully? or I have to interpolate the value...

Here is the sample of data created without taking DeadBallMin(game paused) into account...

** columns A and B are actual data observed during the time-step.

** Exp_Win_A and Exp_Win_B are monotonic increasing. And assume all the features are uniformly distributed within the time-step.

    A      B    Exp_win_A   Exp_win_B   DeadBallMin
0   1      0    0.891713    1.074992    0.000000
1   0      1    0.893859    1.076465    0.000000
2   0      1    0.930300    1.077941    0.036633
3   0      1    0.932539    1.112289    0.000000
4   0      0    0.934783    1.122372    0.907834

From the above table, in the second row, the game was stopped for 33.67% of the time-step.


Any suggestions on how to incorporate the 'DeadBallMin' time into Exp_win_A & Exp_win_B while keeping the behaviour?

  • $\begingroup$ Is high precision estimation favorable for you? $\endgroup$ – Alireza Zolanvari Mar 14 '19 at 11:57
  • $\begingroup$ Yes, I can work with that. $\endgroup$ – Abs Mar 14 '19 at 12:22

One of the most common ways for this purpose is to generate artificial data for that intervals. One of the most novel and powerful algorithms for this purpose is Generative Query Network (GQN).

In this algorithm the basic application is to create a 3D model of an object by observing a few 2D samples of it. So, the network by receiving the angle of the camera beside the 2D image, try to construct the 3D model.

But in this case for time series data, time can play the role of camera angle. So, you can generate high precision artificial data in the case of data lost.

Here is a brief description of this algorithm.

And also here is an implementation of it using Pytorch.

For more information you can read it's respective paper.


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.