2
$\begingroup$

I am developing some classification/regression models form accelerometry time-series data. So far, I have created datapoints by extracting features from non-overlapping sliding windows of the time-series data. I would like to try using overlapping windows as well. However, I was wondering whether it is conceptually sound or there might be some caveats to keep in mind as the data is reused in the overlapping windows.

$\endgroup$

1 Answer 1

2
$\begingroup$

In general I don't see anything wrong with overlapping windows, it might make perfect sense depending on your task. In fact some learning models (e.g. for sequence labeling) do use features based on past data points, which is conceptually similar to having overlaps between them.

However you need to be careful about the fact that this makes data points depend on each other, so of course the preprocessing must be done separately for the training and test set.

$\endgroup$
1
  • $\begingroup$ Thank you. Does it make sense to have different overlap ratios for different classes? $\endgroup$
    – Reza
    Aug 13, 2019 at 6:34

Your Answer

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

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