Investigate Lime suitability for time series

[loostrum]

Potentially interesting: https://github.com/emanuel-metzenthin/Lime-For-Time

[cpranav93]

Investigated Lime-For-Time (https://github.com/emanuel-metzenthin/Lime-For-Time): It has two examples of time-series data-sets:

• coffee (http://timeseriesclassification.com/description.php?Dataset=Coffee)
• ecg (http://timeseriesclassification.com/description.php?Dataset=CinCECGtorso))

The results for the coffee data-set seem interesting and meaningful while for the ecg, I am not sure if it is meaningful (need to look in to it further).

Also, the documentation is limited and there is no publication associated to the work (yet!). Could be interesting to pursue further if relevant.

[elboyran]

After searching for XAI methods for TS (see also issue #345), I found the following:

• Presentation on the topic listing many methods, for post-hoc again GradCAM and LEFTIST.
• Local Explainer For TIme Series classificaTion (LEFTIST) algorithm. It's an adaptation of LIME for TS. Still not clear whether the Lime-For-Time software implements that or imply makes original LIME run for TS. The LEFTIST paper "Agnostic Local Explanation for Time Series Classification".
• Another LIME adaptation LIMESegment for TS was recently published LIMESegment: Meaningful, Realistic Time Series Explanations with software! Seems to also implement 2 other methods (for benchamrking in the paper)- Neves and LEFTIST! Worth investigating?

[loostrum]

We have a separate issues for comparing the different LIME-like implementations, but in any case there are implementations available that looks promising. Closing this issue.