Proformer is a transformer based model for process traces presented in "Structural Positional Encoding for Knowledge Integration in Transformer-based Medical Process Monitoring". Here we present the source code applied to the BPI 2012 challenge dataset.
| BPI2012 | No SPE | SPE |
|---|---|---|
| Acc@1 | 0.8140 | 0.8601 |
| Acc@3 | 0.9742 | 0.9800 |
| Acc@5 | 0.9925 | 0.9939 |
To install the required library using pip:
git clone https://github.com/christopher-irw/proformer_ce.git
cd proformer_ce
pip install -r requirements.txtTo run a training using the best parameters use:
# run on BPI2012 dataset using SPE
python run_proformer.py --use_taxonomy
# run on BPI2012 dataset WITHOUT SPE
python run_proformer.py The notebooks directory contains the dataset preprocessing. In general, it is sufficient to have a csv containing a case_id column representing a unique identifier for the cases and a activity column representing the applied actions.
@article{irwin2024structural,
title={Structural Positional Encoding for knowledge integration in transformer-based medical process monitoring},
author={Irwin, Christopher and Dossena, Marco and Leonardi, Giorgio and Montani, Stefania},
journal={arXiv preprint arXiv:2403.08836},
year={2024}
}