ProcessGAN: Generating Privacy-Preserving Time-Aware Process Data with Conditional Generative Adversarial Nets.

Area of Science:

• Computer Science
• Data Science
• Artificial Intelligence

Background:

• Process data from event logs offers insights into procedural dynamics but is often not sharable due to confidentiality and complexity.
• Limited availability of process data restricts research and analytics in the process mining domain.

Purpose of the Study:

• To address the limitation of sharable process data by introducing a synthetic process data generation method.
• To develop a generative adversarial network (ProcessGAN) capable of creating privacy-preserving process data with realistic activity sequences and timestamps.

Main Methods:

• ProcessGAN utilizes a transformer-based generator and a time-aware self-attention discriminator.
• The model considers process duration and inter-activity time intervals for realistic data generation.
• Evaluated on five real-world datasets (public and private medical), using statistical metrics, supervised model scoring, and domain expert evaluation of discovered workflows.

Main Results:

• ProcessGAN outperforms existing generative models in creating complex processes with parallel pathways.
• Generated synthetic data accurately represents long-range dependencies and authentic timestamp distributions.
• Associated synthetic contexts (e.g., patient demographics) also showed high fidelity compared to authentic data.

Conclusions:

• ProcessGAN effectively generates sharable synthetic process data that is indistinguishable from authentic data.
• The approach enhances the feasibility of research and analytics in process mining, especially for sensitive domains like healthcare.
• The developed model and source code are publicly available to facilitate further research.