not sure if extremely trivial or really tricky.
In the end I want a machine that generates a time series without further input based on training data, generating a new time series every time.
I want this to be based on a transformer.
I want it trained with data looking like this:
2023-07-03 14:19:48,GOOD 2023-07-04 13:59:07,GOOD 2023-07-05 01:58:54,GOOD 2023-07-05 03:30:05,BAD 2023-07-05 05:17:43,BAD 2023-07-06 05:35:34,GOOD 2023-07-07 14:06:03,GOOD 2023-07-08 21:16:05,BAD
with “GOOD” and “BAD” being the state of the system which is likely dependent on the time series data up to that point. I have a lot of data and it’s data points like the one above with maybe a hundred rows of data on average for a few thousand systems. Every system is independent of all others but all are identical.
I do not want to train only on “GOOD” as this would leave out a lot of valueable data …
Is there a way to train a time series transformer with both data that leads to GOOD as well as BAD outcomes, so it would generate time series from scratch which are unlikely to have BAD outcomes?