how to cluster irregularly spaced time series?

[fkiraly]

question from discord helpdesk:

Could anyone direct me to ways of using sktime for clustering if one are dealing with irregularly sampled multivariate time-series. These time-series are only about 30 samples long, and stems from hospital vitals registrations on pulse, temperature, blood pressure, etc. There could be long periods with no registrations, and then periods with high frequency registrations. Moreover, the time series could be of various lengths.

[fkiraly]

Generally, you want to use the all_estimators search utility to search for either clusterers that can deal with irregularly sampled time series, or transformers that turn irregularly sampled time series into regularly sampled ones.
You should be able to pipeline transformers with clusterers that can't deal with irregularly sampled ones to get a clusterer that can deal with irregularly sampled time series

Caveat: the clustering module is still fairly experimental.

The tags to search for:

• "capability:unequal_length" for clusterers
• "capability:unequal_length:removes" for transformers

Generally, you can use the all_tags utility to search for tags and their explanations.

[helgeingvart]

Sorry, didn't see this one until now :-) I have a student that is considering your latest changes #4003 and #3967. I may try to add this functionality to sktime, but regrettably I didn't get many promising results from it so far, but it may relate to poor quality of my data.

[fkiraly]

ah well, you can't always win, I suppose.
If you are up to sharing the data in one of the community sessions, some of the more active community members could have a look. @JonathanBechtel and team are working on clustering and usually attend the Friday meet-ups, so might be interested.

Also, if you have anything for a wishlist, please let us know (e.g., via a wishlist issue!)

[Escape142]

Hi @helgeingvart,

Just a quick question: has the functionality for irregularly sampled time series clustering been added to sktime?

Thanks!

[fkiraly]

as far as I know, no - none of the clusterers support passing of an arbitrary sktime distance, which would allow to construct clusterers ofr unequal length or irregularly sampled time series.

I'm happy to instruct on a mini-project if you would be willing to contribute and/or know someone who would. I estimate 10-20 hours for upgrading one or a few of the existing clusterers, or adding a new one based on an sklearn clusterer.

[Escape142]

What approach do you have in mind to suggest?