Identical shapelets

[Bruno-Hanzen]

When fitting shapelets to a set of time series, I get twice the same shapelet (for the same ts, at different places).
In the ts, there are zones of constant values. 2 of them are 11 positions long, so after normalization, this gives 2 shapelets with 11 "0" in "data", but different "start_pos".

I suggest testing identity or similarity of a new shapelet (data part) with the previously discovered ones. This would save time in the "fit" and "transform" execution.

[mloning]

@jasonlines @TonyBagnall any idea?

[TonyBagnall]

hi, yes this is possible, although we do have a threshold to exclude shapelets with identical info gain. STC could do with some work, its on my list after sorting out distance functions

[Bruno-Hanzen]

Hi, I can confirm that the shapelets were identical, but info gain was different (0.18 and 0.14), if I remember correctly.

[TonyBagnall]

hmmm, curious, we will look into it, thanks for pointing it out

[MatthewMiddlehurst]

The new shapelet_transform in #1490 removes identical shapelets from the final set. This issue still remains for the original transform(s) in shapelets.py, however. I will leave this open for now.

[TonyBagnall]

I think we can close it, original shapelets no longer needed. OP feel free to reopen if you really need the original fixed