Structure-Aware Message Passing for Petri Net Process Mining

New graph neural network architectures, PNRWPE and DA-PNGNN, incorporate directed walks and incidence matrix diffusion to capture cycles and invariants in Petri nets. These methods increased macro balanced accuracy from 0.751 to 0.796 on 2023 Process Discovery Contest models.