Finding the top-k left matrix profile

[colemanyu]

I would like to know whether we can retrieve the top-k left matrix of a time series in a direct way in the current version of stumpy.
I noticed in stumpy.stump, we can obtain the top-k matrix. But how about the top-k left matrix?
I also noticed from #1093, we can find such a top-k left matrix by repeating using stumpy.match, I am curious, are there other ways to do it?

[colemanyu]

I think the easy way to do it is to simply run stumpy.stump for k times.
Each iteration, I will mark the region of the current selected subsequence returned by stumpy.stump, and its nearby region will be the exclusion zone.
Hence, in the next iteration, stumpy.stump would just skip those selected subsequences.

[colemanyu]

I just realized this idea may not work. Because all the subsequences would look at their top-k left neighbors from the SAME time series. I cannot just mark the region as an exclusion only because they are the left neighbors of some subsequences.

[seanlaw]

@colemanyu Thank you for your question and welcome to the STUMPY community.

I would like to know whether we can retrieve the top-k left matrix of a time series in a direct way in the current version of stumpy.

Unfortunately, this is not possible. @NimaSarajpoor might have some thoughts to share as he worked on implementing the top-k feature. As I recall, it was rather complicated/tricky :(

[NimaSarajpoor]

@colemanyu
The answer depends on whether it is acceptable for the top-k neighbors (of a sequence) to lie within each other’s exclusion zones. I’ve provided my answer for both cases below.

(1) The top-k left neighbours (of a subsequence) can be in the each other's exclusion zone

The top-k neighbours (of a subsequence S) returned by stumpy.stump are not in the exclusion zone of S. HOWEVER, those neighbours might be in the exclusion zone of each other. Based on your initial question:

I would like to know whether we can retrieve the top-k left matrix of a time series in a direct way in the current version of stumpy.
I noticed in stumpy.stump, we can obtain the top-k matrix. But how about the top-k left matrix?

It seems that all you want is top-k left neighbours... and you don't care whether those top-k neighbours (of a subsequence S) fall within the exclusion zone of each other. As long as the neighbours are outside of the exclusion zone of S, this is acceptable. In this case, I think there should be a way to modify stump.py to give us the left top-k matrix profile.

(2) The top-k left neighbours should NOT be in each other's exclusion zone.

If this is what you want, then, as mentioned by @seanlaw, it is not possible to achieve this in an efficient way (via stumpy.stump) as far as I know.

---

Alternative approach for small-size data:

You can achieve (1) or (2) by computing the FULL distance matrix (using stumpy.core.mass_distance_matrix), and extract the information you are looking for directly from the distance matrix.