Improving Predictive Process Analytics with Deep Learning and XAI

Obuzor, Princewill

Abstract

In this doctoral thesis, we explore the innovative application of the Tab Transformer
architecture in the realm of predictive process mining, marking a significant advancement in
forecasting subsequent events within activity sequences. Utilising the PM2 methodology,
known for its structured approach in process mining, this study rigorously handles data
processing, model development, and validation. This methodological choice is pivotal in
leveraging the unique capabilities of the Tab Transformer, particularly its proficiency in
processing multiple categorical features, a dimension often overlooked in previous research.
The empirical analysis encompassed a novel dataset and extended to three additional publicly
available datasets: MIMIC-IV Emergency Department (ED) Data, BPIC 2012, BPIC 2013, BPIC
2017, and BPIC Road Traffic. The model's performance was exemplary, achieving accuracies
of 0.69, 0.812, 0.7301, 0.8766, and 0.78, and F1 scores of 0.67, 0.77, 0.70, 0.8533, and 0.734
in these datasets, respectively.
A major contribution of this research is the introduction of the Tab Transformer to process
mining, a first in the field. This approach not only demonstrates the model’s versatility across
various data forms but also highlights the importance of integrating categorical features in
process mining, providing a more nuanced understanding of the influencing factors in activity
sequences.
The thesis further distinguishes itself through the application of Explainable Artificial
Intelligence (XAI) techniques, particularly SHAP and LIME. These tools were instrumental in
demystifying the model’s decision-making processes, thereby enhancing its transparency,
and fostering trust in AI systems. This integration challenges the notion of AI as impenetrable
"black boxes," paving the way for AI systems that are not only effective but also interpretable
and trustworthy.
In conclusion, this thesis contributes significantly to the field of predictive process mining by
pioneering the use of the Tab Transformer, emphasizing the role of categorical features, and
advancing the cause of transparency in AI through the application of XAI. The findings and
methodologies established in this study represent a benchmark for future research in this
evolving domain.

Citation

Obuzor, P. (2024). Improving Predictive Process Analytics with Deep Learning and XAI. (Thesis). University of Salford

• Thesis Type: Thesis
• Deposit Date: Apr 9, 2024
• Publicly Available Date: May 26, 2024
• Award Date: Apr 25, 2024