Contrastive motifs from positive and negative series pairs

[hoangthienan95]

I'm trying to do the following and wonder if it's possible in STUMPY. I have read the tutorials but they don't seem to quite fit my use case perfectly.

Input: Pairs of series (one of type A and one of type B, not equal length). I have 3 classes of pairs. The distribution of each class is below:

• Class 1: ~ 600k pairs
• Class 2: ~ 1k pairs
• Class 3: ~ 1k pairs

I want to find

• multi-dimensional motifs (pattern between both type A and B series) that distinguishes each class (either in pairwise or one vs rest fashion).
• multi-dimensional motifs (pattern between both type A and B series) that are similar.

My use case is different from the Multidimensional_Motif_Discovery tutorial because the motif is repeated in pairs of the same class, not within any series. Also, the motif might not be aligned between type A and B series

For example, the resulting output should be that all Class 1 pairs have the pattern (type A and type B series not matching in sequence but a consistent pair-pattern of length 4)

• Type A series: [2,4,6,8,....]
• Type B series: [2,2,2,2,....]
  while pairs in the other classes dont

And all Class 2 pairs have the pattern (type A and type B series matching sequence of length 4 but at different locations )

• Type A series [1,3,5,7,....]
• Type B series [...1,3,5,7....]
  while pairs in the other classes dont

I hope I didn't make it too confusing, just trying to see if STUMPY could help with my problem!

[seanlaw]

@hoangthienan95 Thank you for your question and welcome to the STUMPY community. So, I fully understand your inputs and outputs but here are a few suggestions:

Since the motifs within your time series pairs are not aligned, you are essentially forced to analyze type A and type B independently for motifs. However, for a given class, I would consider concatenating all type A time series (inserting an np.nan between each series so that no motifs will span independent series) and computing the matrix profile via stumpy.stump(type_A_series, m=4). Then, systematically extract the top n motifs (maybe n = 100). Then, do the same for type B time series. Afterward, do this for the remaining 2 classes. With top n motifs for type A and type B in each class, you can then consider doing a pairwise comparison of the motifs across the 3 classes and see if any class has unique motifs that are not found in the others.

Again, I'm sure I'm not understanding your problem correctly but I believe this should be helpful.