earthgecko
commented
7 years ago @ClimbsRocks the terrible time complexity of current implementations with timeseries is part of the appeal surely ;) Is that not why we are interested? Really it could all be better with timeseries couldn't it? Bizarre that one of the most basic data sets we have is what machine learning is really arguably least good with :) Maybe we are making too complex? However tsfresh does make one aspect so much simpler! And hats off on the shout out at the tsfresh peeps, they are so very nice!
GillesVandewiele
ClimbsRocks
ddofer
There was an interesting paper last year which compared a bunch of newly published time series classifiers.
http://www-bcf.usc.edu/~liu32/milets16/paper/MiLeTS_2016_paper_5.pdf
In a wild oversimplification, several of the classifiers simply transform the data in a new way, then run pretty standard time series classification algorithms on the transformed data.
It would be awesome if tsfresh could become Python's central repository for these different time series data transformations.
The cool part about this project is that they have all the source code available. The team from UEA even reached out to the authors of different papers if the code they produced didn't accurately recreate the results of the published papers.
https://bitbucket.org/TonyBagnall/time-series-classification/src/f4fe66b74e039b0475a87ebf6d57db400da25b63/TimeSeriesClassification/src/tsc_algorithms/COTE.java?at=default&fileviewer=file-view-default
They're very up front about the terrible time complexity of their current implementations. But I think people could probably improve upon that if re-implementing for speed.
This is certainly not an immediate or even near-term project to take on. And I assume that you've already got some of these transforms baked into tsfresh. But I'm interested in trying out some of these techniques in my own work, and figured you guys would probably be interested in some of this stuff too.