0
$\begingroup$

I would like to apply methods like Gramian angular field, recurrence plots and Markov transition fields to a time series classification (TSC) problem where the time series themselves are all of unequal lengths. I wondered if there exist some elegant ways to perform this transform without doing something 'blunt' like zero-padding the time-series.

I did come across the idea of rescaling/resizing post-transformation in https://ai.stackexchange.com/questions/6274/how-can-i-deal-with-images-of-variable-dimensions-when-doing-image-segmentation. They also mention the so-called SPP-net which does look cool : https://arxiv.org/pdf/1406.4729.pdf. However it does require that I want to solve the problem using a deep model (which I might not want to do in the end), and I do see some susceptibility to over-fitting on first glance (especially, if one has a relatively small dataset - which I do). However, this would still require the time series lengths are all the same as input (although I suppose one could consider batching the time series into near-about the same length and padding/downsampling accordingly).

Some preceding points:

  1. One may argue that the easiest thing to do is to look into various summary statistics to generate features and do the TSC from there. I have done this, I do get some results - but I am concerned the these stats are missing something that a visual representation of the problem won't miss.

  2. One may argue that I should just zero-pad (subject to context which in my case; is fine), and go from there. This is probably ok for sequential architectures, but this might corrupt (?) the visual representation. I guess I'm asking if there is a way out of having to do this?

  3. One may insist that I explain the contextual reason for having unequal length time series, this is something I cannot really get into and I strongly do not believe it helps with the technical problem of dealing with image representations of unequal time series. I appreciate that I am being difficult here - if it is impossible to help - I can close the question.

Thanks!

$\endgroup$
0
$\begingroup$

A straightforward approach is to perform the transformation into the visual representation space then use a python tool like PIL or cv2 to rescale the image. For recurrence plots it's not too difficult to recapture the the binary nature of the images but for images like markov transition fields this is a little tougher.

As the asker of this question - I ended up doing this but favoured recurrence plots the most since I had a good idea of how to minimise loss, namely by making the rescaled imagine binary again.

$\endgroup$

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.