tslearn needs uniformly sampled and same sized time series

[MaxBenChrist]

After a first glimpse on the package, I was wondering if tslearn needs uniformly sampled and same sized time series?

[rtavenar]

Hi,

tslearn accepts time series of different lengths in a dataset (see here for example), but there is a clear lack in the documentation (and the checks done in the code) about whether each algorithm expects equal sized time series or not.

Concerning sampling, at the moment you are right, there is a uniform sampling assumption (mainly because the algorithms implemented for now do not care about sampling if I am correct. This is also something that should be discussed at some point: if you have experience on that, I would be glad to hear/read you on that point.

[rtavenar]

In the end, this issue seems related to Issue #2 . So I might close this one as soon as time series format is properly documented and (most importantly) associated machine learning algorithms get flags in the docs indicating if:

1. they don't care about the time series format (i.e. are supposed to run whatever the format, which should be tested) ;
2. they accept only time series of equal length.

[MaxBenChrist]

There are four main questions that arise when working with time series data:

1. Do the time series can have different lengths?
2. Are the time series uniformly sampled?
3. Are the time series allowed to have missing values, NaNs etc.?
4. Are multivariate time series allowed?

Depending on the answers, matrix based formats can not be used, instead one has to use stacked formats.

I think the bast place to have a discussion on this topic would be in my repository with the list of time series python packages. I will add some documents and then link it to here.

[MaxBenChrist]

See

MaxBenChrist/awesome_time_series_in_python#1

and

https://github.com/MaxBenChrist/awesome_time_series_in_python/blob/master/standardize_time_series_formats.md

regarding the format discussion

[rtavenar]

A note has been added to the doc of all methods that do not accept time series of different lengths. So if nothing specified, the algo should run even if the dataset has time series of different lengths.