This repository provides materials to reproduce the results presented in the paper "Remaining cycle time prediction with Graph Neural Networks for Predictive Process Monitoring".
The 1_Data folder contains raw data of Helpdesk and BPIC20 datasets. We can not provide the EMS3141 data due to the confidentiality reasons. The statistics of theses logs are shown in the table below:
| Event log | Num. cases | Num. activities | Num. events | Avg. case length | Max. case lenth | Avg. case duration (days) | Max. case duration (days) | Min. case duration (days) | Variants |
|---|---|---|---|---|---|---|---|---|---|
| Helpdesk | 4552 | 10 | 21197 | 4.66 | 15 | 40.85 | 59.99 | 30.64 | 207 |
| BPIC20 | 10043 | 15 | 55130 | 5.49 | 24 | 11.62 | 368.19 | 1.06 | 64 |
| EMS3141 | 99282 | 35 | 287117 | 28.76 | 44 | 7.01 | 80.87 | 0.69 | 296 |
The project is implemented with jupyter notebook in both python and pytorch.
- pytorch: Follow the instructions in the PyTorch website https://pytorch.org/get-started/locally/
- torch geometric for GNN implementation: https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html
- Ax package for hyperparamaters tuning:
pip3 install ax-platform - PM4Py:
pip install pm4py
There are 4 .ipynb files in 3_Notebooks folder. The notebook Data_processing.ipynb must be run first to obtain the processed data. Then, Prefix_encoding_GGNN.ipynb is to prepare data for GGNN model. Two last notebooks each correspond to a model, i.e., LSTM and Gated GNN.