implement lime for time series

[cwmeijer]

LimeSegment

assume convenient input

including tests

See #369 and #366 for background.

[cwmeijer]

LimeSegment doesn't have a pypi package so installation isn't the nicest. Looking at their code, the actual LimeSegment implementation seems both simple and tiny, similar to RISE. We should consider creating our own implementation based on theirs. This would give us more flexibility and 1 dependency less to worry about. See https://github.com/TortySivill/LIMESegment/blob/0a276e30f8d259642521407e7d51d07969169432/Utils/explanations.py#L94

[cwmeijer]

Loading and running the coffee model was not working because I made some error while writing the coffee.onnx a few months ago. I fixed both the script, and the model in the surf drive.

[cwmeijer]

Some problems with LimeSegment:

• I can't get any sensible output for the coffee dataset; result is all zeros.
• There is no license provided with the LimeSegment code
• LimeSegment doesn't have any setup.cfg or pypi registration

We'll now try Lime-for-Time first. Blocked by #453.

[cpranav93]

Update: I managed to get limesegment running for the coffee data-set but with a caveat. It works for an output prediction of 1 but not for 0. This is related to the RIDGE linear regression model and the fact that the model we have for a coffee type (y = 0) predicts consistently y = 0 even when parts of the time-series are masked. Thus for the regression equation Y - wx: Y is always 0 leading to a trivial solution of w = 0. I hope this makes sense.

Conclusion: Stick with lime for time for now as limesegment has some major flaws for basic data-sets.