0
$\begingroup$

Say I have time series data for classifying stars using deep learning based on stellar variability, with each time series data measuring the flux of the star overtime. For each star, I have the data split into an equal number of bins, but the time between the first bin and the last bin is not consistent. For example:

Star A Star B
T : F T : F
-------- --------
0.5 : 3 0.2 : 2
1.0 : 5 0.4 : 6
1.5 : 1 0.6 : 7

Am I better off standardising the time columns to a consistent difference across all my data for the deep learning models to learn? E.g should I change it to this:

Star A Star B
T : F T : F
-------- --------
1.0 : 3 1.0 : 2
2.0 : 5 2.0 : 6
3.0 : 1 3.0 : 7

I'm assuming the models need a standard input size but now I'm not sure thinking about it

$\endgroup$

1 Answer 1

1
$\begingroup$

Most time series classification algorithms do not take the time index as input, so there's no point standardising the time index. The description for some algorithms will mention a time index, but actually use the position in the sequence, rather than a supplied index.

For these algorithms I would create all the time series to use consistent time steps. If the algorithm handles variable length time series, then use all the data. If the algorithm requires equal length time series, then extract equal length sub-sequences from each time series. This obviously also assumes that the time span of the shortest time series is long enough to capture the flux of all the stars.

If you are using an algorithm that takes a supplied time index, then I'd use the non-standardised time index as in your top figure.

$\endgroup$

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.