Efficient Training of Recurrent Neural Networks for Remaining Time Prediction in Predictive Process Monitoring

Roider J, Zanca D, Eskofier B (2024)

Publication Language: English

Publication Type: Conference contribution, Original article

Publication year: 2024

Publisher: Springer

Series: BPM: International Conference on Business Process Management

City/Town: Cham

Book Volume: 22

Pages Range: 238 - 255

Conference Proceedings Title: Business Process Management, 22nd International Conference, BPM 2024, Krakow, Poland, September 1–6, 2024, Proceedings

Event location: Krakow PL

ISBN: 978-3-031-70395-9

URI: https://link.springer.com/book/10.1007/978-3-031-70396-6

DOI: 10.1007/978-3-031-70396-6

Abstract

An important task in predictive process monitoring is the prediction of the remaining time. The accuracy of methods to solve this task has been steadily improved and verified during the last years. The current state-of-the-art uses Long Short-Term Memory (LSTM) neural networks, represented by the DA-LSTM model. However, training such methods requires substantial amounts of time and memory. This paper addresses specifically these problems. We adjust DA-LSTM and introduce DA-LSTM+ which achieves competitive error levels while reducing the training time significantly. Furthermore, we introduce trace-based sequence encoding as an alternative to prefix encoding, and an approach to use case attributes more efficiently to address time and memory limitations during training. The usage of these is not limited to LSTM’s but they are compatible with any neural network type. We evaluate them together with two alternatives for categorical feature encoding in an extensive benchmark study including eight different model architectures based on DA-LSTM+ and 14 publicly available datasets. The study shows that the training of neural network-based methods can be significantly accelerated with our contributions without affecting model’s performance. Our implementation is memory-efficient and publicly available.

Authors with CRIS profile

Johannes Roider Machine Learning and Data Analytics Lab Dario Zanca Machine Learning and Data Analytics Lab Björn Eskofier Machine Learning and Data Analytics Lab

Additional Organisation(s)

Institute Materials for Electronics and Energy Technology (i-MEET)

How to cite

APA:

Roider, J., Zanca, D., & Eskofier, B. (2024). Efficient Training of Recurrent Neural Networks for Remaining Time Prediction in Predictive Process Monitoring. In Andrea Marrella, Manuel Resinas, Mieke Jans, Michael Rosemann (Eds.), Business Process Management, 22nd International Conference, BPM 2024, Krakow, Poland, September 1–6, 2024, Proceedings (pp. 238 - 255). Krakow, PL: Cham: Springer.

MLA:

Roider, Johannes, Dario Zanca, and Björn Eskofier. "Efficient Training of Recurrent Neural Networks for Remaining Time Prediction in Predictive Process Monitoring." Proceedings of the BPM: International Conference on Business Process Management, Krakow Ed. Andrea Marrella, Manuel Resinas, Mieke Jans, Michael Rosemann, Cham: Springer, 2024. 238 - 255.

BibTeX: Download