Dealing with unequal length and irregular time series in time series classification, regression, forecasting, clustering, etc

[fkiraly]

As this is a frequently asked question, I'm summarizing an answer to this.
(mid-term, we probably want to put this in the tutorials)

There are two main ways to deal with unequal length or irregularly spaced time series:

• search for an estimator of the right type, e.g., classifier, clusterer, that can deal with irregularly spaced time series. The tag in question is capability:unequal_length - despite the name, afaik current estimators of this type can deal with both unequal length and irregular spacing. To search, use registry.all_estimators (and see docstring).
• pipelining with a transformer that renders series equally spaced. The tag for this is capability:unequal_length:removes, search with registry.all_estimators. To pipeline, use pipeline.make_pipeline (and see docstring). Eamples would be TimeBinAggregate from series.binning, or TSInterpolator from panel.interpolate.

If the above does not work as implied, please raise a bug report here:
https://github.com/sktime/sktime/issues/new?assignees=&labels=bug&template=bug_report.md&title=%5BBUG%5D

[anonymous-engineering]

Also have the same question - any update on time series forecasters

[fkiraly]

@anonymouspapi, can you kindly explain what you mean - what precisely is your question?

[jonaslandsgesell]

Is it or will it be possible to make use of numpy masked arrays https://numpy.org/doc/stable/reference/maskedarray.generic.html to pass time series of unequal length to sktime?

I imagine to pass a numpy (masked) array X with X.shape = [num_time_series, num_features, max_length_time_series]
Now not all time series would have to be of length max_lenghr_time_series. The time series which are shorter would be filled with a mask value (fill value) so that the subsequent pipelines do not make inappropriate use of these fill values.

Alternatively: is there a conversion function for sich numpy masked arrays to a suitable pandas data frame (if that exists)?

[fkiraly]

@jonaslandsgesell, apologies, did not see this.

I was not aware of masked arrays, I think this is an excellent suggestion!

For conversion, you would use datatypes.convert or convert_to, not sure though whether that works for masked arrays.