Multi-PMF: Multivariate Approaches for Process Model Forecasting

A framework for time series forecasting of business process models. This project combines process model forecasting techniques with advanced machine learning and deep learning models to predict process behavior.

Overview

This project provides tools for:

• Preprocessing event logs
• Generating daily process model (Directly-Follows Graph, DFG)
• Converting process model to time series data
• Training and evaluating multiple forecasting models

Project Structure

.
├── event_logs/           # Raw and processed event logs
├── data/                 # Daily DFGs and activity mapping
├── dataset/              # Processed time series datasets
├── lib/                  # Library files
├── preprocess.py         # Event log preprocessing
├── gen_daily_dfg.py      # Generates daily DFG matrics
├── dfg_to_time_series.py # Converts DFGs to DF time series
└── multiDFpred.py        # Main prediction script

Features

• Event Log Preprocessing

  • Filters infrequent process variants
  • Trims event logs to specified time ranges
  • Adds artificial start/end activities

• DF Time Series Generation

  • Extracts DFGs with activity mapping
  • Time series conversion

• Multiple Forecasting Models

  • Baseline models:
    • Persistence
    • Mean
  • Advanced models:
    • Linear Regression
    • XGBoost
    • LightGBM
    • RNN/LSTM/GRU
    • N-BEATS
    • N-HiTS
    • TCN
    • Transformer
    • TFT
    • DLinear
    • NLinear

Requirements

• Python 3.9
• PyTorch
• Darts
• PM4Py
• NumPy
• Pandas
• Matplotlib
• Scikit-learn
• XGBoost
• LightGBM
• h5py

Usage

1. Preprocess Event Logs

python preprocess.py

2. Generate Daily DFGs

python gen_daily_dfg.py

3. Convert to DF Time Series

python dfg_to_time_series.py

4. Run Predictions

python multiDFpred.py --dataset <dataset_name>

Dataset Setup

1. Download the event logs from the following sources:

   • BPI2019_1 - BPI Challenge 2019
   • Hospital_Billing - Hospital Billing Event Log
   • RTFMP - Road Traffic Fine Management Process

2. Create the required directory structure:

   mkdir -p event_logs/original

3. Place the downloaded .xes files in the event_logs/original directory

4. After running python preprocess.py, the processed event logs will be saved in the ./event_logs/processed directory.

Model Training

The framework select the following parameters:

• Input sequence length: 32 (days)
• Output sequence length: 16 (days)
• Training/Test split: 80/20

Evaluation Metrics

• Mean Absolute Error (MAE)
• Root Mean Square Error (RMSE)

Repository

For the complete repository for PMF benchmark including additional resources, documentation, and examples, please visit: PMF-Benchmark

Reference

De Smedt, J., Yeshchenko, A., Polyvyanyy, A., De Weerdt, J., & Mendling, J. (2023). Process model forecasting and change exploration using time series analysis of event sequence data. Data & Knowledge Engineering, 145, 102145.

Citation

If you find this repository helpful for your work, please consider citing our paper:

@inproceedings{yu2024multivariate,
  title={Multivariate Approaches for Process Model Forecasting},
  author={Yu, Yongbo and Peeperkorn, Jari and De Smedt, Johannes and De Weerdt, Jochen},
  booktitle={International Conference on Process Mining},
  pages={279--292},
  year={2024},
  organization={Springer}
}