Request: stumpi equivalent for multivariate case
theconnectionist opened this issue · comments
Is there a stumpi equivalent for use with multiple variables (multiple timeseries) so I can incrementally add a new record? I read the article on using stumpi (much appreciated). When I initially asked chatGPT about incrementally updating a multivariate matrix profile in stumpy, it helpfully suggested "mstumpi". Looking at the documentation, it appears ChatGPT had a hallucination and sent me searching for that function :-)
Right now it appears the only way to do it is to batch a few streaming events and recompute the matrix in its entirety, perhaps using mstumped to benefit from distributed computing.
Unfortunately, there is no mstumpi equivalent as, fundamentally, multi-dimensional matrix profiles have a different/specific definition that is different from simply stacking individual matrix profiles one on top of the other. The interpretation of a multi-dimensional matrix profile is not so obvious. See a past mstumpi discussion here.
Are you really looking for a multi-dimensional matrix profile (i.e., the same as the one that is produced by mstump)?
@seanlaw, I read through and played with mstumped to generate and visualize multi-dimensional timeseries. I understand the limitation you mentioned, but taking time to let that "multiple dimensions are not multiple single dimensions stacked together" sink in. I'll read what you pointed me to.
@theconnectionist In case you missed it, I strongly recommend reading through this tutorial for multi-dimensional matrix profiles as it goes through the (originally published) concept in detail.