Predicting Remaining Cycle Time from Ongoing Cases: A Survival Analysis-Based Approach / Fadi Nader Zaki Baskharon

Supervisor: Mohamed A. ElHelw

Thesis (M.A.)—Nile University, Egypt, 2021 .

"Includes bibliographical references"

Contents:
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Thesis Outline and Summary of Contributions . . . . . . . . . . . . . . . . . . 4
2. Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.2 Survival Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.3 Weibull distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.4 Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Baseline Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3. Proposed Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1 Model Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Optimization function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Output interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4. Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.1 Experiment I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.2 Experiment II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
vii
5. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1 Future directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Appendices:
A. cycle prediction documentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Bibliography .

Abstract:
in predictive process monitoring. Different approaches that learn from event logs, e.g., relying
on an existing representation of the process or leveraging machine learning approaches, have
been proposed in the literature to tackle this problem. Machine learning-based techniques have
shown superiority over other techniques with respect to the accuracy of the prediction as well
as freedom from knowledge about the underlying process models generating the logs. However,
all proposed approaches only learn from complete traces. This might cause delays in starting
new training cycles as usually process instances last over a long time that could even reach
months or years.
In this thesis, we propose a machine learning approach that can also accept and learn from
incomplete (ongoing) traces. Using a time-aware survival analysis technique, we can train a
neural network to predict the most likely remaining cycle time of a running case.
This approach is evaluated on different real-life datasets and is compared with a state-ofthe-
art baseline. Results show that our approach, in most cases, is able to outperform the
baseline approach with a simple model architecture and less training time.
The approach is further enhanced to learn from trace level - fixed - features as well as the
events-related features. We empirically proved that trace-level features enhance the prediction
power of the model using a real-life dataset.

Text in English, abstracts in English.