Add parameter T_subseq_isfinite to stumpy.motifs

[NimaSarajpoor]

I was wondering if we could add T_subseq_isfinite to the stumpy.motifs. It can then be used in the following block of code in function _motifs:

query_matches = match(
            Q,
            T,
            M_T=M_T,
            Σ_T=Σ_T,
            max_matches=None,
            max_distance=max_distance,
            atol=atol,
        )

[seanlaw]

@ninimama I'm not sure what you mean? z-normalized functions like stumpy.motifs don't use T_subseq_isfinite anywhere. Only non-normalized methods use T_subseq_isfinite. What you please explain what you are you trying to accomplish?

[NimaSarajpoor]

z-normalized functions like stumpy.motifs don't use T_subseq_isfinite anywhere. Only non-normalized methods use T_subseq_isfinite

Didn't know that. My case is for non-normalized distance though. I was trying to find each motif (and their matches) and then exclude them and their trivial matches, and then iterate through the process (i.e. finding the next motif) till I cover the whole time series.

I know that the function _aamp_motifs has a similar concept:

for idx in query_matches[:, 1]:
            core.apply_exclusion_zone(P, int(idx), excl_zone, np.inf)

However, if I understand correctly, we here apply exclusion on the matrix profile. What I am looking for is something different. I want to apply it on distance profile for query matching, which can be done by T_subseq_isfinite. I might make a branch and experience it later. I understand it might not be good to have one-for-all purpose function. So, feel free to close this if you think it is better to keep the function as is.

[seanlaw]

If you are dealing with query matching, have you looked at the stumpy.match function? It's a bit hard for me to understand what it is that you want to accomplish without more information.

[NimaSarajpoor]

Please let me know if this explanation helps.

This is what I am trying to achieve: Given a time series T (and there is no query Q), I want to get the very-top motif and their top-k matches. Then, I want to get the next motif. However, I want to completely ignore the previously-discovered motif (i.e. I do not want it to affect the matrix profile P in the first place. In other words, I want to set their corresponding indices to np.inf in distance profile). This way, I can make sure that after some iterations, my motifs (and their matches) will exhaust the whole time series T. I acknowledge that as I go through the iterations based on the aforementioned approach, my later motifs might get a very large distance to their NN, which is from the remaining indices. But, I am okay with that.

---

This can be seen as some sort of clustering. The general idea is to group the top-motif with its k_NN matches; and then go through the rest of subsequences. (I know that we are not optimizing any objective function here and I know there are some overlaps between subsequences; but I just want to see IF I FORCE my time series T to be broken down to some similar subsequence, how would it look like?)

[seanlaw]

Yeah, I think this is a rather specialized (i.e., not general) use case and beyond the scope of those functions. So, you'll have to write your own code to handle this. Of course, I am here to help! 😄