Open shfa5275 opened 4 years ago
Thank you for your suggestion!
I have a question. The multimatrixprofile gives $d$ matrixprofiles. Taking the max on each matrixprofile gives $d$ discord locations, which can be too much. I am wondering that: what I first take the mean or the median of all the matrixprofiles and then compute the max? Is it a sufficient approximation of a discord location in an overall sense?
Thanks again and kind regards!
Well, you have to try and see if the results make sense for you. Motif make sense to aggregate (i.e.: having a motif that is similar across dimensions. For discord, you could try the overall discord of the discord for each dimension.
I think MultiMatrixProfile aims to have a meaningful motif search algorithm (as in the paper). Meanwhile, you can try to search the largest value in the computed profile. This for sure is a nice feature to add. I'll mark this as a TODO.