Generating a synthetic training data source for ML-based process mining tools

Anjali SINGH

Master's thesis

Abstract:

This thesis addresses the issue of data scarcity in process mining by utilizing Generative Adversarial Networks (GANs) to generate synthetic training data. Process mining tools, which rely on machine learning, often face challenges due to the limited availability of high-quality data, which is crucial for their effectiveness. The study's key steps include selecting and preprocessing authentic process event logs, developing and training a specialized GAN model to produce synthetic data, and evaluating the quality of this data using statistical metrics. A comparative analysis is then performed to assess the performance of the GAN model against a Long Short-Term Memory (LSTM) model, using both synthetic and real datasets (BPI Challenge 2012 and Helpdesk). Results indicate that while LSTM models accurately replicate the original data structure, GANs introduce greater variability, providing a wider range of training scenarios. This suggests that GAN-generated data could enhance process mining tools by exposing them to more diverse and realistic process patterns.

Abstract:

This thesis addresses the issue of data scarcity in process mining by utilizing Generative Adversarial Networks (GANs) to generate synthetic training data. Process mining tools, which rely on machine learning, often face challenges due to the limited availability of high-quality data, which is crucial for their effectiveness. The study's key steps include selecting and preprocessing authentic process event logs, developing and training a specialized GAN model to produce synthetic data, and evaluating the quality of this data using statistical metrics. A comparative analysis is then performed to assess the performance of the GAN model against a Long Short-Term Memory (LSTM) model, using both synthetic and real datasets (BPI Challenge 2012 and Helpdesk). Results indicate that while LSTM models accurately replicate the original data structure, GANs introduce greater variability, providing a wider range of training scenarios. This suggests that GAN-generated data could enhance process mining tools by exposing them to more diverse and realistic process patterns.

Keywords

Process Mining, Data Scarcity, Generative Adversarial Networks (GANs), Synthetic Training Data, Machine Learning, Transformers, Sequence Generation, Event Logs, LSTM Model, Data Preprocessing, Comparative Analysis

Language used: English

Date on which the thesis was submitted / produced: 20. 8. 2024

Thesis defence

• Supervisor: prof. Dr. Andreas Fischer

UNIVERSITY OF SOUTH BOHEMIA IN ČESKÉ BUDĚJOVICE

Faculty of Science

Master programme / field:
Artificial Intelligence and Data Science / Artificial Intelligence and Data Science