I'd love to build this data into a neural network and see if I can come up with robosongs. It should be quite good at coming up with "What chord should come next?"
So if you open up a book like 'Music Theory for Complete Idiots' there's a section on "Chord progressions" that essentially has what you'd recognize as a state transition table for "what chords will sound good after this chord?" and I'm pretty sure there's a row that looks like [iii => IV, vi].
Coming up with a computerized vocabulary for the elements of coherent large-scale composition structures would be more of an interesting area to research than individual chord transitions, because the latter is really a solved problem.
A friend of mine has a site that does something similar. It uses music theory to generate real-sounding music for musicians to practice sight-reading.
https://sightreadingfactory.com/
Coming up with a computerized vocabulary for the elements of coherent large-scale composition structures would be more of an interesting area to research than individual chord transitions, because the latter is really a solved problem.