For this repo and paper: https://github.com/diogoff/unlabelled-event-logs
Business processes are modeled as Markov and Expectation Maximization is used to find the model.
So suppose a business process has states A,B,C,D,E and I see
A,C,D A,B, A,C,D,E
That repo will give me the probabilities of seeing each as well as the table describing the Markov model.
Now my company is using word2vec to predict what state comes in between two states.
Then I thought of attention from this website I read a while back: https://jalammar.github.io/
If I feed transformers enough sequences that are actually seen, will it be able to predict next states?
If so, what specific downstream task can this be classified as?
I am looking at models on Hugging Face and have no idea where to start.